DADHosp: HOSPITAL.COST regression:: DADHosp_SLR_bodywt
bdanalytics
Date: (Tue) Dec 08, 2015
Introduction:
Data: Source: Training: DADHosp_Cost.txt
New:
Time period:
Synopsis:
Based on analysis utilizing <> techniques,
Summary of key steps & error improvement stats:
Prediction Accuracy Enhancement Options:
- transform.data chunk:
- derive features from multiple features
- manage.missing.data chunk:
- Not fill missing vars
- Fill missing numerics with a different algorithm
- Fill missing chars with data based on clusters
- extract.features chunk:
- Text variables: move to date extraction chunk ???
- Mine acronyms
- Mine places
- Text variables: move to date extraction chunk ???
- Review set_global_options chunk after features are finalized
[](.png)
Potential next steps include:
- Organization:
- Categorize by chunk
- Priority criteria:
- Ease of change
- Impacts report
- Cleans innards
- Bug report
- all chunks:
- at chunk-end rm(!glb_)
- manage.missing.data chunk:
- cleaner way to manage re-splitting of training vs. new entity
- extract.features chunk:
- Add n-grams for glbFeatsText
- “RTextTools”, “tau”, “RWeka”, and “textcat” packages
- Convert user-specified mutate code to config specs
- Add n-grams for glbFeatsText
- fit.models chunk:
- Prediction accuracy scatter graph:
- Add tiles (raw vs. PCA)
- Use shiny for drop-down of “important” features
Use plot.ly for interactive plots ?
- Change .fit suffix of model metrics to .mdl if it’s data independent (e.g. AIC, Adj.R.Squared - is it truly data independent ?, etc.)
- create a custom model for rpart that has minbucket as a tuning parameter
varImp for randomForest crashes in caret version:6.0.41 -> submit bug report
- Probability handling for multinomials vs. desired binomial outcome
- ROCR currently supports only evaluation of binary classification tasks (version 1.0.7)
extensions toward multiclass classification are scheduled for the next release
Skip trControl.method=“cv” for dummy classifier ?
- fit.all.training chunk:
- myplot_prediction_classification: displays ‘x’ instead of ‘+’ when there are no prediction errors
- Compare glb_sel_mdl vs. glb_fin_mdl:
- varImp
- Prediction differences (shd be minimal ?)
- Move glb_analytics_diag_plots to mydsutils.R: (+) Easier to debug (-) Too many glb vars used
- Add print(ggplot.petrinet(glb_analytics_pn) + coord_flip()) at the end of every major chunk
- Parameterize glb_analytics_pn
- Move glb_impute_missing_data to mydsutils.R: (-) Too many glb vars used; glb_<>_df reassigned
- Do non-glm methods handle interaction terms ?
- f-score computation for classifiers should be summation across outcomes (not just the desired one ?)
- Add accuracy computation to glb_dmy_mdl in predict.data.new chunk
- Why does splitting fit.data.training.all chunk into separate chunks add an overhead of ~30 secs ? It’s not rbind b/c other chunks have lower elapsed time. Is it the number of plots ?
- Incorporate code chunks in print_sessionInfo
- Test against
- projects in github.com/bdanalytics
- lectures in jhu-datascience track
Analysis:
rm(list = ls())
set.seed(12345)
options(stringsAsFactors = FALSE)
source("~/Dropbox/datascience/R/myscript.R")
source("~/Dropbox/datascience/R/mydsutils.R")## Loading required package: caret
## Loading required package: lattice
## Loading required package: ggplot2source("~/Dropbox/datascience/R/myplot.R")
source("~/Dropbox/datascience/R/mypetrinet.R")
source("~/Dropbox/datascience/R/myplclust.R")
source("~/Dropbox/datascience/R/mytm.R")
# Gather all package requirements here
suppressPackageStartupMessages(require(doMC))
registerDoMC(6) # # of cores on machine - 2
suppressPackageStartupMessages(require(caret))
require(plyr)## Loading required package: plyrrequire(dplyr)## Loading required package: dplyr
##
## Attaching package: 'dplyr'
##
## The following objects are masked from 'package:plyr':
##
## arrange, count, desc, failwith, id, mutate, rename, summarise,
## summarize
##
## The following objects are masked from 'package:stats':
##
## filter, lag
##
## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, union#source("dbgcaret.R")
#packageVersion("snow")
#require(sos); findFn("cosine", maxPages=2, sortby="MaxScore")
# Analysis control global variables
# Inputs
# url/name = "<obsTrnFileName>"; sep = choose from c(NULL, "\t")
glbObsTrnFile <- list(url = NULL, name = "DADHosp_Cost.txt", sep = "\t")
glbObsNewFileName <- "<obsNewFileName>"
glbInpMerge <- NULL #: default
# list(fnames = c("<fname1>", "<fname2>")) # files will be concatenated
glb_is_separate_newobs_dataset <- FALSE # or TRUE
glb_split_entity_newobs_datasets <- TRUE # FALSE not supported - use "copy" for glb_split_newdata_method # select from c(FALSE, TRUE)
glb_split_newdata_method <- "copy" # select from c(NULL, "condition", "sample", "copy")
glb_split_newdata_condition <- NULL # or "is.na(<var>)"; "<var> <condition_operator> <value>"
glb_split_newdata_size_ratio <- 0.3 # > 0 & < 1
glb_split_sample.seed <- 123 # or any integer
glbObsDropCondition <- NULL # : default
# "<condition>" # use | & ; NOT || &&
#parse(text=glbObsDropCondition)
#subset(glbObsAll, .grpid %in% c(31))
glb_obs_repartition_train_condition <- NULL # : default
# "<condition>"
glb_max_fitobs <- NULL # or any integer
glb_is_regression <- TRUE; glb_is_classification <- !glb_is_regression;
glb_is_binomial <- NULL # or TRUE or FALSE
glb_rsp_var_raw <- "HOSPITAL.COST"
# for classification, the response variable has to be a factor
glb_rsp_var <- glb_rsp_var_raw # or "HOSPITAL.COST.fctr"
# if the response factor is based on numbers/logicals e.g (0/1 OR TRUE/FALSE vs. "A"/"B"),
# or contains spaces (e.g. "Not in Labor Force")
# caret predict(..., type="prob") crashes
glb_map_rsp_raw_to_var <- NULL
# function(raw) {
# return(raw ^ 0.5)
# return(log(raw))
# return(log(1 + raw))
# return(log10(raw))
# return(exp(-raw / 2))
# ret_vals <- rep_len(NA, length(raw)); ret_vals[!is.na(raw)] <- ifelse(raw[!is.na(raw)] == 1, "Y", "N"); return(relevel(as.factor(ret_vals), ref="N"))
# #as.factor(paste0("B", raw))
# #as.factor(gsub(" ", "\\.", raw))
# }
# glb_map_rsp_raw_to_var(tst <- c(NA, 0, 1))
# glb_map_rsp_raw_to_var(tst <- c(NA, 0, 2.99, 280.50, 1000.00))
glb_map_rsp_var_to_raw <- NULL
# function(var) {
# return(var ^ 2.0)
# return(exp(var))
# return(10 ^ var)
# return(-log(var) * 2)
# as.numeric(var)
# gsub("\\.", " ", levels(var)[as.numeric(var)])
# c("<=50K", " >50K")[as.numeric(var)]
# c(FALSE, TRUE)[as.numeric(var)]
# }
# glb_map_rsp_var_to_raw(glb_map_rsp_raw_to_var(tst))
if ((glb_rsp_var != glb_rsp_var_raw) && is.null(glb_map_rsp_raw_to_var))
stop("glb_map_rsp_raw_to_var function expected")
# List info gathered for various columns
# <col_name>: <description>; <notes>
# If multiple vars are parts of id, consider concatenating them to create one id var
# If glb_id_var == NULL, ".rownames <- as.numeric(row.names())" is the default
# User-specified exclusions
glbFeatsExclude <- c(NULL
# Feats that shd be excluded due to known causation by prediction variable
# , "<feat1", "<feat2>"
,"HospCost.cut.fctr"
# Feats that are linear combinations (alias in glm)
# Feature-engineering phase -> start by excluding all features except id & category & work each one in
# ,"BODY.WEIGHT"
)
if (glb_rsp_var_raw != glb_rsp_var)
glbFeatsExclude <- union(glbFeatsExclude, glb_rsp_var_raw)
glbFeatsInteractionOnly <- list()
#glbFeatsInteractionOnly[["<child_feat>"]] <- "<parent_feat>"
# currently does not handle more than 1 column; consider concatenating multiple columns
glb_id_var <- "PTID" # choose from c(NULL : default, "<id_feat>")
glbFeatsCategory <- "HospCost.cut.fctr" # choose from c(NULL : default, "<category_feat>")
glb_drop_vars <- c(NULL
# , "<feat1>", "<feat2>"
)
glb_map_vars <- NULL # or c("<var1>", "<var2>")
glb_map_urls <- list();
# glb_map_urls[["<var1>"]] <- "<var1.url>"
glb_assign_pairs_lst <- NULL;
# glb_assign_pairs_lst[["<var1>"]] <- list(from=c(NA),
# to=c("NA.my"))
glb_assign_vars <- names(glb_assign_pairs_lst)
# Derived features; Use this mechanism to cleanse data ??? Cons: Data duplication ???
glbFeatsDerive <- list();
# glbFeatsDerive[["<feat.my.sfx>"]] <- list(
# mapfn = function(<arg1>, <arg2>) { return(function(<arg1>, <arg2>)) }
# , args = c("<arg1>", "<arg2>"))
# character
# mapfn = function(Week) { return(substr(Week, 1, 10)) }
# mapfn = function(descriptor) { return(plyr::revalue(descriptor, c(
# "ABANDONED BUILDING" = "OTHER",
# "**" = "**"
# ))) }
# mapfn = function(description) { mod_raw <- description;
# This is here because it does not work if it's in txt_map_filename
# mod_raw <- gsub(paste0(c("\n", "\211", "\235", "\317", "\333"), collapse = "|"), " ", mod_raw)
# Don't parse for "." because of ".com"; use customized gsub for that text
# mod_raw <- gsub("(\\w)(!|\\*|,|-|/)(\\w)", "\\1\\2 \\3", mod_raw);
# return(mod_raw) }
#print(mod_raw <- grep(""", glbObsAll[, txt_var], value = TRUE))
#print(mod_raw <- glbObsAll[c(88,187,280,1040,1098), txt_var])
#print(mod_raw <- glbObsAll[sel_obs(list(descr.my.contains="\\bdoes( +)not\\b")), glbFeatsText])
#print(mod_raw <- glbObsAll[sel_obs(list(descr.my.contains="\\bipad [[:digit:]]\\b")), glbFeatsText][01:10])
#print(mod_raw <- glbObsAll[sel_obs(list(descr.my.contains="pad mini")), glbFeatsText][11:20])
#print(mod_raw <- glbObsAll[sel_obs(list(descr.my.contains="pad mini")), glbFeatsText][21:30])
#print(mod_raw <- glbObsAll[sel_obs(list(descr.my.contains="pad mini")), glbFeatsText][31:40])
#glbObsAll[which(glb_post_stop_words_terms_mtrx_lst[[txt_var]][, subset(glb_post_stop_words_terms_df_lst[[txt_var]], term %in% c("conditionminimal"))$pos] > 0), "description"]
# numeric
# Create feature based on record position/id in data
glbFeatsDerive[[".pos"]] <- list(
mapfn = function(.rnorm) { return(1:length(.rnorm)) }
, args = c(".rnorm"))
glbFeatsDerive[["HospCost.cut.fctr"]] <- list(
mapfn = function(HOSPITAL.COST) { return(cut(HOSPITAL.COST, 5, breaks = c(0, 100000, 200000, 300000, 900000), labels = NULL)) }
, args = c("HOSPITAL.COST"))
# Add logs of numerics that are not distributed normally
# Derive & keep multiple transformations of the same feature, if normality is hard to achieve with just one transformation
# Right skew: logp1; sqrt; ^ 1/3; logp1(logp1); log10; exp(-<feat>/constant)
# glbFeatsDerive[["WordCount.log1p"]] <- list(
# mapfn = function(WordCount) { return(log1p(WordCount)) }
# , args = c("WordCount"))
# glbFeatsDerive[["WordCount.root2"]] <- list(
# mapfn = function(WordCount) { return(WordCount ^ (1/2)) }
# , args = c("WordCount"))
# glbFeatsDerive[["WordCount.nexp"]] <- list(
# mapfn = function(WordCount) { return(exp(-WordCount)) }
# , args = c("WordCount"))
#print(summary(glbObsAll$WordCount))
#print(summary(mapfn(glbObsAll$WordCount)))
# mapfn = function(HOSPI.COST) { return(cut(HOSPI.COST, 5, breaks = c(0, 100000, 200000, 300000, 900000), labels = NULL)) }
# mapfn = function(Rasmussen) { return(ifelse(sign(Rasmussen) >= 0, 1, 0)) }
# mapfn = function(startprice) { return(startprice ^ (1/2)) }
# mapfn = function(startprice) { return(log(startprice)) }
# mapfn = function(startprice) { return(exp(-startprice / 20)) }
# mapfn = function(startprice) { return(scale(log(startprice))) }
# mapfn = function(startprice) { return(sign(sprice.predict.diff) * (abs(sprice.predict.diff) ^ (1/10))) }
# factor
# mapfn = function(PropR) { return(as.factor(ifelse(PropR >= 0.5, "Y", "N"))) }
# mapfn = function(productline, description) { as.factor(gsub(" ", "", productline)) }
# mapfn = function(purpose) { return(relevel(as.factor(purpose), ref="all_other")) }
# mapfn = function(raw) { tfr_raw <- as.character(cut(raw, 5));
# tfr_raw[is.na(tfr_raw)] <- "NA.my";
# return(as.factor(tfr_raw)) }
# mapfn = function(startprice.log10) { return(cut(startprice.log10, 3)) }
# mapfn = function(startprice.log10) { return(cut(sprice.predict.diff, c(-1000, -100, -10, -1, 0, 1, 10, 100, 1000))) }
# , args = c("<arg1>"))
# multiple args
# mapfn = function(PTS, oppPTS) { return(PTS - oppPTS) }
# mapfn = function(startprice.log10.predict, startprice) {
# return(spdiff <- (10 ^ startprice.log10.predict) - startprice) }
# mapfn = function(productline, description) { as.factor(
# paste(gsub(" ", "", productline), as.numeric(nchar(description) > 0), sep = "*")) }
# # If glbObsAll is not sorted in the desired manner
# mapfn=function(Week) { return(coredata(lag(zoo(orderBy(~Week, glbObsAll)$ILI), -2, na.pad=TRUE))) }
# mapfn=function(ILI) { return(coredata(lag(zoo(ILI), -2, na.pad=TRUE))) }
# mapfn=function(ILI.2.lag) { return(log(ILI.2.lag)) }
# glbFeatsDerive[["<var1>"]] <- glbFeatsDerive[["<var2>"]]
glb_derive_vars <- names(glbFeatsDerive)
# tst <- "descr.my"; args_lst <- NULL; for (arg in glbFeatsDerive[[tst]]$args) args_lst[[arg]] <- glbObsAll[, arg]; print(head(args_lst[[arg]])); print(head(drv_vals <- do.call(glbFeatsDerive[[tst]]$mapfn, args_lst)));
# print(which_ix <- which(args_lst[[arg]] == 0.75)); print(drv_vals[which_ix]);
glbFeatsDateTime <- list()
# glbFeatsDateTime[["<DateTimeFeat>"]] <-
# c(format = "%Y-%m-%d %H:%M:%S", timezone = "America/New_York", impute.na = TRUE,
# last.ctg = TRUE, poly.ctg = TRUE)
glbFeatsPrice <- NULL # or c("<price_var>")
glbFeatsText <- NULL # c("<txt_var>") # NULL #
Sys.setlocale("LC_ALL", "C") # For english## [1] "C/C/C/C/C/en_US.UTF-8"# Text Processing Step: custom modifications not present in txt_munge -> use glbFeatsDerive
# Text Processing Step: universal modifications
glb_txt_munge_filenames_pfx <- "<projectId>_mytxt_"
# Text Processing Step: tolower
# Text Processing Step: myreplacePunctuation
# Text Processing Step: removeWords
glb_txt_stop_words <- list()
# Remember to use unstemmed words
if (!is.null(glbFeatsText)) {
require(tm)
glb_txt_stop_words[["<txt_var>"]] <- sort(c(NULL
# Remove any words from stopwords
# , setdiff(myreplacePunctuation(stopwords("english")), c("<keep_wrd1>", <keep_wrd2>"))
# cor.y.train == NA
# ,unlist(strsplit(paste(c(NULL
# ,"<comma-separated-terms>"
# ), collapse=",")
# freq == 1; keep c("<comma-separated-terms-to-keep>")
# ,<comma-separated-terms>
# chisq.pval high (e.g. == 1); keep c("<comma-separated-terms-to-keep>")
# ,<comma-separated-terms>
# nzv.freqRatio high (e.g. >= glb_nzv_freqCut); keep c("<comma-separated-terms-to-keep>")
# ,<comma-separated-terms>
))
}
#orderBy(~term, glb_post_stem_words_terms_df_lst[[txt_var]][grep("^2", glb_post_stem_words_terms_df_lst[[txt_var]]$term), ])
#glbObsAll[glb_post_stem_words_terms_mtrx_lst[[txt_var]][, 6] > 0, glbFeatsText]
# To identify terms with a specific freq
#paste0(sort(subset(glb_post_stop_words_terms_df_lst[[txt_var]], freq == 1)$term), collapse = ",")
#paste0(sort(subset(glb_post_stem_words_terms_df_lst[[txt_var]], freq <= 2)$term), collapse = ",")
# To identify terms with a specific freq &
# are not stemmed together later OR is value of color.fctr (e.g. gold)
#paste0(sort(subset(glb_post_stop_words_terms_df_lst[[txt_var]], (freq == 1) & !(term %in% c("blacked","blemish","blocked","blocks","buying","cables","careful","carefully","changed","changing","chargers","cleanly","cleared","connect","connects","connected","contains","cosmetics","default","defaulting","defective","definitely","describe","described","devices","displays","drop","drops","engravement","excellant","excellently","feels","fix","flawlessly","frame","framing","gentle","gold","guarantee","guarantees","handled","handling","having","install","iphone","iphones","keeped","keeps","known","lights","line","lining","liquid","liquidation","looking","lots","manuals","manufacture","minis","most","mostly","network","networks","noted","opening","operated","performance","performs","person","personalized","photograph","physically","placed","places","powering","pre","previously","products","protection","purchasing","returned","rotate","rotation","running","sales","second","seconds","shipped","shuts","sides","skin","skinned","sticker","storing","thats","theres","touching","unusable","update","updates","upgrade","weeks","wrapped","verified","verify") ))$term), collapse = ",")
#print(subset(glb_post_stem_words_terms_df_lst[[txt_var]], (freq <= 2)))
#glbObsAll[which(terms_mtrx[, 229] > 0), glbFeatsText]
# To identify terms with cor.y == NA
#orderBy(~-freq+term, subset(glb_post_stop_words_terms_df_lst[[txt_var]], is.na(cor.y)))
#paste(sort(subset(glb_post_stop_words_terms_df_lst[[txt_var]], is.na(cor.y))[, "term"]), collapse=",")
#orderBy(~-freq+term, subset(glb_post_stem_words_terms_df_lst[[txt_var]], is.na(cor.y)))
# To identify terms with low cor.y.abs
#head(orderBy(~cor.y.abs+freq+term, subset(glb_post_stem_words_terms_df_lst[[txt_var]], !is.na(cor.y))), 5)
# To identify terms with high chisq.pval
#subset(glb_post_stem_words_terms_df_lst[[txt_var]], chisq.pval > 0.99)
#paste0(sort(subset(glb_post_stem_words_terms_df_lst[[txt_var]], (chisq.pval > 0.99) & (freq <= 10))$term), collapse=",")
#paste0(sort(subset(glb_post_stem_words_terms_df_lst[[txt_var]], (chisq.pval > 0.9))$term), collapse=",")
#head(orderBy(~-chisq.pval+freq+term, glb_post_stem_words_terms_df_lst[[txt_var]]), 5)
#glbObsAll[glb_post_stem_words_terms_mtrx_lst[[txt_var]][, 68] > 0, glbFeatsText]
#orderBy(~term, glb_post_stem_words_terms_df_lst[[txt_var]][grep("^m", glb_post_stem_words_terms_df_lst[[txt_var]]$term), ])
# To identify terms with high nzv.freqRatio
#summary(glb_post_stem_words_terms_df_lst[[txt_var]]$nzv.freqRatio)
#paste0(sort(setdiff(subset(glb_post_stem_words_terms_df_lst[[txt_var]], (nzv.freqRatio >= glb_nzv_freqCut) & (freq < 10) & (chisq.pval >= 0.05))$term, c( "128gb","3g","4g","gold","ipad1","ipad3","ipad4","ipadair2","ipadmini2","manufactur","spacegray","sprint","tmobil","verizon","wifion"))), collapse=",")
# To identify obs with a txt term
#tail(orderBy(~-freq+term, glb_post_stop_words_terms_df_lst[[txt_var]]), 20)
#mydspObs(list(descr.my.contains="non"), cols=c("color", "carrier", "cellular", "storage"))
#grep("ever", dimnames(terms_stop_mtrx)$Terms)
#which(terms_stop_mtrx[, grep("ipad", dimnames(terms_stop_mtrx)$Terms)] > 0)
#glbObsAll[which(terms_stop_mtrx[, grep("16", dimnames(terms_stop_mtrx)$Terms)[1]] > 0), c(glbFeatsCategory, "storage", txt_var)]
# To identify whether terms shd be synonyms
#orderBy(~term, glb_post_stop_words_terms_df_lst[[txt_var]][grep("^moder", glb_post_stop_words_terms_df_lst[[txt_var]]$term), ])
# term_row_df <- glb_post_stop_words_terms_df_lst[[txt_var]][grep("^came$", glb_post_stop_words_terms_df_lst[[txt_var]]$term), ]
#
# cor(glb_post_stop_words_terms_mtrx_lst[[txt_var]][glbObsAll$.lcn == "Fit", term_row_df$pos], glbObsTrn[, glb_rsp_var], use="pairwise.complete.obs")
# To identify which stopped words are "close" to a txt term
#sort(cluster_vars)
# Text Processing Step: stemDocument
# To identify stemmed txt terms
#glb_post_stop_words_terms_df_lst[[txt_var]][grep("condit", glb_post_stop_words_terms_df_lst[[txt_var]]$term), ]
#orderBy(~term, glb_post_stem_words_terms_df_lst[[txt_var]][grep("^con", glb_post_stem_words_terms_df_lst[[txt_var]]$term), ])
#glbObsAll[which(terms_stem_mtrx[, grep("use", dimnames(terms_stem_mtrx)$Terms)[[1]]] > 0), c(glb_id_var, "productline", txt_var)]
#glbObsAll[which(TfIdf_stem_mtrx[, 191] > 0), c(glb_id_var, glbFeatsCategory, txt_var)]
#which(glbObsAll$UniqueID %in% c(11915, 11926, 12198))
# Text Processing Step: mycombineSynonyms
# To identify which terms are associated with not -> combine "could not" & "couldn't"
#findAssocs(glb_full_DTM_lst[[txt_var]], "not", 0.05)
# To identify which synonyms should be combined
#orderBy(~term, glb_post_stem_words_terms_df_lst[[txt_var]][grep("^c", glb_post_stem_words_terms_df_lst[[txt_var]]$term), ])
chk_comb_cor <- function(syn_lst) {
# cor(terms_stem_mtrx[glbObsAll$.src == "Train", grep("^(damag|dent|ding)$", dimnames(terms_stem_mtrx)[[2]])], glbObsTrn[, glb_rsp_var], use="pairwise.complete.obs")
print(subset(glb_post_stem_words_terms_df_lst[[txt_var]], term %in% syn_lst$syns))
print(subset(get_corpus_terms(tm_map(glb_txt_corpus_lst[[txt_var]], mycombineSynonyms, list(syn_lst), lazy=FALSE)), term == syn_lst$word))
# cor(terms_stop_mtrx[glbObsAll$.src == "Train", grep("^(damage|dent|ding)$", dimnames(terms_stop_mtrx)[[2]])], glbObsTrn[, glb_rsp_var], use="pairwise.complete.obs")
# cor(rowSums(terms_stop_mtrx[glbObsAll$.src == "Train", grep("^(damage|dent|ding)$", dimnames(terms_stop_mtrx)[[2]])]), glbObsTrn[, glb_rsp_var], use="pairwise.complete.obs")
}
#chk_comb_cor(syn_lst=list(word="cabl", syns=c("cabl", "cord")))
#chk_comb_cor(syn_lst=list(word="damag", syns=c("damag", "dent", "ding")))
#chk_comb_cor(syn_lst=list(word="dent", syns=c("dent", "ding")))
#chk_comb_cor(syn_lst=list(word="use", syns=c("use", "usag")))
glb_txt_synonyms <- list()
#glb_txt_synonyms[["<txt_var>"]] <- list(NULL
# , list(word="<stem1>", syns=c("<stem1>", "<stem1_2>"))
# )
# options include: "weightTf", "myweightTflog1p", "myweightTfsqrt", "weightTfIdf", "weightBM25"
glb_txt_terms_control <- list(weighting = "weightTfIdf" # : default
# termFreq selection criteria across obs: tm default: list(global=c(1, Inf))
, bounds = list(global = c(1, Inf))
# wordLengths selection criteria: tm default: c(3, Inf)
, wordLengths = c(1, Inf)
)
glb_txt_cor_var <- glb_rsp_var # : default # or c(<feat>)
# select one from c("union.top.val.cor", "top.cor", "top.val", default: "top.chisq", "sparse")
glbFeatsTextFilter <- "top.chisq"
glbFeatsTextTermsMax <- rep(10, length(glbFeatsText)) # :default
names(glbFeatsTextTermsMax) <- glbFeatsText
# Text Processing Step: extractAssoc
glbFeatsTextAssocCor <- rep(1, length(glbFeatsText)) # :default
names(glbFeatsTextAssocCor) <- glbFeatsText
# Remember to use stemmed terms
glb_important_terms <- list()
# Text Processing Step: extractPatterns (ngrams)
glbFeatsTextPatterns <- list()
#glbFeatsTextPatterns[[<txt_var>>]] <- list()
#glbFeatsTextPatterns[[<txt_var>>]] <- c(metropolitan.diary.colon = "Metropolitan Diary:")
# Have to set it even if it is not used
# Properties:
# numrows(glb_feats_df) << numrows(glbObsFit
# Select terms that appear in at least 0.2 * O(FP/FN(glbObsOOB)) ???
# numrows(glbObsOOB) = 1.1 * numrows(glbObsNew) ???
glb_sprs_thresholds <- NULL # or c(<txt_var1> = 0.988, <txt_var2> = 0.970, <txt_var3> = 0.970)
glbFctrMaxUniqVals <- 20 # default: 20
glb_impute_na_data <- FALSE # or TRUE
glb_mice_complete.seed <- 144 # or any integer
glb_cluster <- FALSE # : default or TRUE
glb_cluster.seed <- 189 # or any integer
glb_cluster_entropy_var <- NULL # c(glb_rsp_var, as.factor(cut(glb_rsp_var, 3)), default: NULL)
glbFeatsTextClusterVarsExclude <- FALSE # default FALSE
glb_interaction_only_feats <- NULL # : default or c(<parent_feat> = "<child_feat>")
glb_nzv_freqCut <- 19 # 19 : caret default
glb_nzv_uniqueCut <- 10 # 10 : caret default
glbRFESizes <- list()
#glbRFESizes[["mdlFamily"]] <- c(4, 8, 16, 32, 64, 67, 68, 69) # Accuracy@69/70 = 0.8258
glbObsFitOutliers <- list()
# If outliers.n >= 10; consider concatenation of interaction vars
# glbObsFitOutliers[["<mdlFamily>"]] <- c(NULL
# is.na(.rstudent)
# is.na(.dffits)
# .hatvalues >= 0.99
# -38,167,642 < minmax(.rstudent) < 49,649,823
# , <comma-separated-<glb_id_var>>
# )
glbObsTrnOutliers <- list()
# influence.measures: car::outlier; rstudent; dffits; hatvalues; dfbeta; dfbetas
#mdlId <- "RFE.X.glm"; obs_df <- fitobs_df
#mdlId <- "Final.glm"; obs_df <- trnobs_df
#mdlId <- "CSM2.X.glm"; obs_df <- fitobs_df
#print(outliers <- car::outlierTest(glb_models_lst[[mdlId]]$finalModel))
#mdlIdFamily <- paste0(head(unlist(str_split(mdlId, "\\.")), -1), collapse="."); obs_df <- dplyr::filter_(obs_df, interp(~(!(var %in% glbObsFitOutliers[[mdlIdFamily]])), var = as.name(glb_id_var))); model_diags_df <- cbind(obs_df, data.frame(.rstudent=stats::rstudent(glb_models_lst[[mdlId]]$finalModel)), data.frame(.dffits=stats::dffits(glb_models_lst[[mdlId]]$finalModel)), data.frame(.hatvalues=stats::hatvalues(glb_models_lst[[mdlId]]$finalModel)));print(summary(model_diags_df[, c(".rstudent",".dffits",".hatvalues")])); table(cut(model_diags_df$.hatvalues, breaks=c(0.00, 0.98, 0.99, 1.00)))
#print(subset(model_diags_df, is.na(.rstudent))[, glb_id_var])
#print(subset(model_diags_df, is.na(.dffits))[, glb_id_var])
#print(model_diags_df[which.min(model_diags_df$.dffits), ])
#print(subset(model_diags_df, .hatvalues > 0.99)[, glb_id_var])
#dffits_df <- merge(dffits_df, outliers_df, by="row.names", all.x=TRUE); row.names(dffits_df) <- dffits_df$Row.names; dffits_df <- subset(dffits_df, select=-Row.names)
#dffits_df <- merge(dffits_df, glbObsFit, by="row.names", all.x=TRUE); row.names(dffits_df) <- dffits_df$Row.names; dffits_df <- subset(dffits_df, select=-Row.names)
#subset(dffits_df, !is.na(.Bonf.p))
#mdlId <- "CSM.X.glm"; vars <- myextract_actual_feats(row.names(orderBy(reformulate(c("-", paste0(mdlId, ".imp"))), myget_feats_imp(glb_models_lst[[mdlId]]))));
#model_diags_df <- glb_get_predictions(model_diags_df, mdlId, glb_rsp_var)
#obs_ix <- row.names(model_diags_df) %in% names(outliers$rstudent)[1]
#obs_ix <- which(is.na(model_diags_df$.rstudent))
#obs_ix <- which(is.na(model_diags_df$.dffits))
#myplot_parcoord(obs_df=model_diags_df[, c(glb_id_var, glbFeatsCategory, ".rstudent", ".dffits", ".hatvalues", glb_rsp_var, paste0(glb_rsp_var, mdlId), vars[1:min(20, length(vars))])], obs_ix=obs_ix, id_var=glb_id_var, category_var=glbFeatsCategory)
#model_diags_df[row.names(model_diags_df) %in% names(outliers$rstudent)[c(1:2)], ]
#ctgry_diags_df <- model_diags_df[model_diags_df[, glbFeatsCategory] %in% c("Unknown#0"), ]
#myplot_parcoord(obs_df=ctgry_diags_df[, c(glb_id_var, glbFeatsCategory, ".rstudent", ".dffits", ".hatvalues", glb_rsp_var, "startprice.log10.predict.RFE.X.glmnet", indep_vars[1:20])], obs_ix=row.names(ctgry_diags_df) %in% names(outliers$rstudent)[1], id_var=glb_id_var, category_var=glbFeatsCategory)
#table(glbObsFit[model_diags_df[, glbFeatsCategory] %in% c("iPad1#1"), "startprice.log10.cut.fctr"])
#glbObsFit[model_diags_df[, glbFeatsCategory] %in% c("iPad1#1"), c(glb_id_var, "startprice")]
# No outliers & .dffits == NaN
#myplot_parcoord(obs_df=model_diags_df[, c(glb_id_var, glbFeatsCategory, glb_rsp_var, "startprice.log10.predict.RFE.X.glmnet", indep_vars[1:10])], obs_ix=seq(1:nrow(model_diags_df))[is.na(model_diags_df$.dffits)], id_var=glb_id_var, category_var=glbFeatsCategory)
# Modify mdlId to (build & extract) "<FamilyId>#<Fit|Trn>#<caretMethod>#<preProc1.preProc2>#<samplingMethod>"
glb_models_lst <- list(); glb_models_df <- data.frame()
# Regression
if (glb_is_regression) {
glbMdlMethods <- c(NULL
# deterministic
#, "lm", # same as glm
, "glm", "bayesglm", "glmnet"
, "rpart"
# non-deterministic
, "gbm", "rf"
# Unknown
, "nnet" , "avNNet" # runs 25 models per cv sample for tunelength=5
, "svmLinear", "svmLinear2"
, "svmPoly" # runs 75 models per cv sample for tunelength=5
, "svmRadial"
, "earth"
, "bagEarth" # Takes a long time
)
} else
# Classification - Add ada (auto feature selection)
if (glb_is_binomial)
glbMdlMethods <- c(NULL
# deterministic
, "bagEarth" # Takes a long time
, "glm", "bayesglm", "glmnet"
, "nnet"
, "rpart"
# non-deterministic
, "gbm"
, "avNNet" # runs 25 models per cv sample for tunelength=5
, "rf"
# Unknown
, "lda", "lda2"
# svm models crash when predict is called -> internal to kernlab it should call predict without .outcome
, "svmLinear", "svmLinear2"
, "svmPoly" # runs 75 models per cv sample for tunelength=5
, "svmRadial"
, "earth"
) else
glbMdlMethods <- c(NULL
# non-deterministic
, "rf"
# Unknown
, "gbm", "rpart"
)
glbMdlFamilies <- list(); glb_mdl_feats_lst <- list()
# family: Choose from c("RFE.X", "CSM.X", "All.X", "Best.Interact")
# methods: Choose from c(NULL, <method>, glbMdlMethods)
#glbMdlFamilies[["RFE.X"]] <- c("glmnet", "glm") # non-NULL vector is mandatory
glbMdlFamilies[["All.X"]] <- "glmnet" # non-NULL vector is mandatory
#glbMdlFamilies[["Best.Interact"]] <- "glmnet" # non-NULL vector is mandatory
# Check if interaction features make RFE better
# glbMdlFamilies[["CSM.X"]] <- setdiff(glbMdlMethods, c("lda", "lda2")) # crashing due to category:.clusterid ??? #c("glmnet", "glm") # non-NULL list is mandatory
# glb_mdl_feats_lst[["CSM.X"]] <- c(NULL
# , <comma-separated-features-vector>
# )
# dAFeats.CSM.X %<d-% c(NULL
# # Interaction feats up to varImp(RFE.X.glmnet) >= 50
# , <comma-separated-features-vector>
# , setdiff(myextract_actual_feats(predictors(rfe_fit_results)), c(NULL
# , <comma-separated-features-vector>
# ))
# )
# glb_mdl_feats_lst[["CSM.X"]] <- "%<d-% dAFeats.CSM.X"
glbMdlFamilies[["Final"]] <- c(NULL) # NULL vector acceptable
glbMdlAllowParallel <- list()
#glbMdlAllowParallel[["<mdlId>"]] <- FALSE
# Check if tuning parameters make fit better; make it mdlFamily customizable ?
glbMdlTuneParams <- data.frame()
# When glmnet crashes at model$grid with error: ???
glmnetTuneParams <- rbind(data.frame()
,data.frame(parameter = "alpha", vals = "0.100 0.325 0.550 0.775 1.000")
,data.frame(parameter = "lambda", vals = "9.342e-02")
)
glbMdlTuneParams <- myrbind_df(glbMdlTuneParams,
cbind(data.frame(mdlId = "<mdlId>"),
glmnetTuneParams))
#avNNet
# size=[1] 3 5 7 9; decay=[0] 1e-04 0.001 0.01 0.1; bag=[FALSE]; RMSE=1.3300906
#bagEarth
# degree=1 [2] 3; nprune=64 128 256 512 [1024]; RMSE=0.6486663 (up)
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "bagEarth", parameter = "nprune", vals = "256")
# ,data.frame(method = "bagEarth", parameter = "degree", vals = "2")
# ))
#earth
# degree=[1]; nprune=2 [9] 17 25 33; RMSE=0.1334478
#gbm
# shrinkage=0.05 [0.10] 0.15 0.20 0.25; n.trees=100 150 200 [250] 300; interaction.depth=[1] 2 3 4 5; n.minobsinnode=[10]; RMSE=0.2008313
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "gbm", parameter = "shrinkage", min = 0.05, max = 0.25, by = 0.05)
# ,data.frame(method = "gbm", parameter = "n.trees", min = 100, max = 300, by = 50)
# ,data.frame(method = "gbm", parameter = "interaction.depth", min = 1, max = 5, by = 1)
# ,data.frame(method = "gbm", parameter = "n.minobsinnode", min = 10, max = 10, by = 10)
# #seq(from=0.05, to=0.25, by=0.05)
# ))
#glmnet
# alpha=0.100 [0.325] 0.550 0.775 1.000; lambda=0.0005232693 0.0024288010 0.0112734954 [0.0523269304] 0.2428800957; RMSE=0.6164891
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "glmnet", parameter = "alpha", vals = "0.550 0.775 0.8875 0.94375 1.000")
# ,data.frame(method = "glmnet", parameter = "lambda", vals = "9.858855e-05 0.0001971771 0.0009152152 0.0042480525 0.0197177130")
# ))
#nnet
# size=3 5 [7] 9 11; decay=0.0001 0.001 0.01 [0.1] 0.2; RMSE=0.9287422
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "nnet", parameter = "size", vals = "3 5 7 9 11")
# ,data.frame(method = "nnet", parameter = "decay", vals = "0.0001 0.0010 0.0100 0.1000 0.2000")
# ))
#rf # Don't bother; results are not deterministic
# mtry=2 35 68 [101] 134; RMSE=0.1339974
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "rf", parameter = "mtry", vals = "2 5 9 13 17")
# ))
#rpart
# cp=0.020 [0.025] 0.030 0.035 0.040; RMSE=0.1770237
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "rpart", parameter = "cp", vals = "0.004347826 0.008695652 0.017391304 0.021739130 0.034782609")
# ))
#svmLinear
# C=0.01 0.05 [0.10] 0.50 1.00 2.00 3.00 4.00; RMSE=0.1271318; 0.1296718
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "svmLinear", parameter = "C", vals = "0.01 0.05 0.1 0.5 1")
# ))
#svmLinear2
# cost=0.0625 0.1250 [0.25] 0.50 1.00; RMSE=0.1276354
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method = "svmLinear2", parameter = "cost", vals = "0.0625 0.125 0.25 0.5 1")
# ))
#svmPoly
# degree=[1] 2 3 4 5; scale=0.01 0.05 [0.1] 0.5 1; C=0.50 1.00 [2.00] 3.00 4.00; RMSE=0.1276130
# glbMdlTuneParams <- myrbind_df(glbMdlTuneParams, rbind(data.frame()
# ,data.frame(method="svmPoly", parameter="degree", min=1, max=5, by=1) #seq(1, 5, 1)
# ,data.frame(method="svmPoly", parameter="scale", vals="0.01, 0.05, 0.1, 0.5, 1")
# ,data.frame(method="svmPoly", parameter="C", vals="0.50, 1.00, 2.00, 3.00, 4.00")
# ))
#svmRadial
# sigma=[0.08674323]; C=0.25 0.50 1.00 [2.00] 4.00; RMSE=0.1614957
#glb2Sav(); all.equal(sav_models_df, glb_models_df)
glb_preproc_methods <- NULL
# c("YeoJohnson", "center.scale", "range", "pca", "ica", "spatialSign")
# Baseline prediction model feature(s)
glb_Baseline_mdl_var <- NULL # or c("<feat>")
glbMdlMetric_terms <- NULL # or matrix(c(
# 0,1,2,3,4,
# 2,0,1,2,3,
# 4,2,0,1,2,
# 6,4,2,0,1,
# 8,6,4,2,0
# ), byrow=TRUE, nrow=5)
glbMdlMetricSummary <- NULL # or "<metric_name>"
glbMdlMetricMaximize <- NULL # or FALSE (TRUE is not the default for both classification & regression)
glbMdlMetricSummaryFn <- NULL # or function(data, lev=NULL, model=NULL) {
# confusion_mtrx <- t(as.matrix(confusionMatrix(data$pred, data$obs)))
# #print(confusion_mtrx)
# #print(confusion_mtrx * glbMdlMetric_terms)
# metric <- sum(confusion_mtrx * glbMdlMetric_terms) / nrow(data)
# names(metric) <- glbMdlMetricSummary
# return(metric)
# }
glbMdlCheckRcv <- FALSE # Turn it on when needed; otherwise takes long time
glb_rcv_n_folds <- 3 # or NULL
glb_rcv_n_repeats <- 3 # or NULL
glb_clf_proba_threshold <- NULL # 0.5
# Model selection criteria
if (glb_is_regression)
glbMdlMetricsEval <- c("min.RMSE.OOB", "max.R.sq.OOB", "max.Adj.R.sq.fit", "min.RMSE.fit")
#glbMdlMetricsEval <- c("min.RMSE.fit", "max.R.sq.fit", "max.Adj.R.sq.fit")
if (glb_is_classification) {
if (glb_is_binomial)
glbMdlMetricsEval <-
c("max.Accuracy.OOB", "max.AUCROCR.OOB", "max.AUCpROC.OOB", "min.aic.fit", "max.Accuracy.fit") else
glbMdlMetricsEval <- c("max.Accuracy.OOB", "max.Kappa.OOB")
}
# select from NULL [no ensemble models], "auto" [all models better than MFO or Baseline], c(mdl_ids in glb_models_lst) [Typically top-rated models in auto]
glb_mdl_ensemble <- NULL
# "%<d-% setdiff(mygetEnsembleAutoMdlIds(), 'CSM.X.rf')"
# c(<comma-separated-mdlIds>
# )
# Only for classifications; for regressions remove "(.*)\\.prob" form the regex
# tmp_fitobs_df <- glbObsFit[, grep(paste0("^", gsub(".", "\\.", mygetPredictIds$value, fixed = TRUE), "CSM\\.X\\.(.*)\\.prob"), names(glbObsFit), value = TRUE)]; cor_mtrx <- cor(tmp_fitobs_df); cor_vctr <- sort(cor_mtrx[row.names(orderBy(~-Overall, varImp(glb_models_lst[["Ensemble.repeatedcv.glmnet"]])$imp))[1], ]); summary(cor_vctr); cor_vctr
#ntv.glm <- glm(reformulate(indep_vars, glb_rsp_var), family = "binomial", data = glbObsFit)
#step.glm <- step(ntv.glm)
glb_sel_mdl_id <- "All.X##rcv#glmnet" #select from c(NULL, "All.X##rcv#glmnet", "RFE.X##rcv#glmnet", <mdlId>)
glb_fin_mdl_id <- NULL #select from c(NULL, glb_sel_mdl_id)
glb_dsp_cols <- c(glb_id_var, glbFeatsCategory, glb_rsp_var
# List critical cols excl. glb_id_var, glbFeatsCategory & glb_rsp_var
)
# Output specs
glbOutDataVizFname <- NULL # choose from c(NULL, "<projectId>_obsall.csv")
glb_out_obs <- NULL # select from c(NULL, "all", "new", "trn")
glb_out_vars_lst <- list()
# glb_id_var will be the first output column, by default
if (glb_is_classification && glb_is_binomial) {
glb_out_vars_lst[["Probability1"]] <-
"%<d-% mygetPredictIds(glb_rsp_var, glb_fin_mdl_id)$prob"
} else {
glb_out_vars_lst[[glb_rsp_var]] <-
"%<d-% mygetPredictIds(glb_rsp_var, glb_fin_mdl_id)$value"
}
# glb_out_vars_lst[[glb_rsp_var_raw]] <- glb_rsp_var_raw
# glb_out_vars_lst[[paste0(head(unlist(strsplit(mygetPredictIds$value, "")), -1), collapse = "")]] <-
glbOutStackFnames <- NULL #: default
# c("ebayipads_txt_assoc1_out_bid1_stack.csv") # manual stack
# c("ebayipads_finmdl_bid1_out_nnet_1.csv") # universal stack
glb_out_pfx <- "DADHosp_SLR_bodywt_"
glb_save_envir <- FALSE # or TRUE
# Depict process
glb_analytics_pn <- petrinet(name = "glb_analytics_pn",
trans_df = data.frame(id = 1:6,
name = c("data.training.all","data.new",
"model.selected","model.final",
"data.training.all.prediction","data.new.prediction"),
x=c( -5,-5,-15,-25,-25,-35),
y=c( -5, 5, 0, 0, -5, 5)
),
places_df=data.frame(id=1:4,
name=c("bgn","fit.data.training.all","predict.data.new","end"),
x=c( -0, -20, -30, -40),
y=c( 0, 0, 0, 0),
M0=c( 3, 0, 0, 0)
),
arcs_df=data.frame(
begin=c("bgn","bgn","bgn",
"data.training.all","model.selected","fit.data.training.all",
"fit.data.training.all","model.final",
"data.new","predict.data.new",
"data.training.all.prediction","data.new.prediction"),
end =c("data.training.all","data.new","model.selected",
"fit.data.training.all","fit.data.training.all","model.final",
"data.training.all.prediction","predict.data.new",
"predict.data.new","data.new.prediction",
"end","end")
))
#print(ggplot.petrinet(glb_analytics_pn))
print(ggplot.petrinet(glb_analytics_pn) + coord_flip())## Loading required package: grid
glb_analytics_avl_objs <- NULL
glb_chunks_df <- myadd_chunk(NULL, "import.data")## label step_major step_minor label_minor bgn end elapsed
## 1 import.data 1 0 0 10.55 NA NAStep 1.0: import data
chunk option: eval=
## [1] "Reading file ./data/DADHosp_Cost.txt..."
## [1] "dimensions of data in ./data/DADHosp_Cost.txt: 248 rows x 3 cols"
## PTID BODY.WEIGHT HOSPITAL.COST
## 1 1 49 660293
## 2 2 41 809130
## 3 3 47 362231
## 4 4 80 629990
## 5 5 58 444876
## 6 6 45 372357
## PTID BODY.WEIGHT HOSPITAL.COST
## 9 9 72 437529.1
## 38 38 59 260036.0
## 127 127 18 145362.0
## 179 179 10 180728.0
## 180 180 55 144134.0
## 244 244 69 295155.0
## PTID BODY.WEIGHT HOSPITAL.COST
## 243 243 62 73682
## 244 244 69 295155
## 245 245 57 200321
## 246 246 58 191188
## 247 247 65 202807
## 248 248 71 248112
## 'data.frame': 248 obs. of 3 variables:
## $ PTID : int 1 2 3 4 5 6 7 8 9 10 ...
## $ BODY.WEIGHT : int 49 41 47 80 58 45 60 44 72 77 ...
## $ HOSPITAL.COST: num 660293 809130 362231 629990 444876 ...
## - attr(*, "comment")= chr "glbObsTrn"
## NULL## PTID BODY.WEIGHT HOSPITAL.COST
## 1 1 49 660293
## 2 2 41 809130
## 3 3 47 362231
## 4 4 80 629990
## 5 5 58 444876
## 6 6 45 372357
## PTID BODY.WEIGHT HOSPITAL.COST
## 1 1 49 660293.0
## 44 44 65 178100.0
## 96 96 58 143278.8
## 97 97 56 214679.0
## 99 99 59 262582.0
## 114 114 5 178398.0
## PTID BODY.WEIGHT HOSPITAL.COST
## 243 243 62 73682
## 244 244 69 295155
## 245 245 57 200321
## 246 246 58 191188
## 247 247 65 202807
## 248 248 71 248112
## 'data.frame': 248 obs. of 3 variables:
## $ PTID : int 1 2 3 4 5 6 7 8 9 10 ...
## $ BODY.WEIGHT : int 49 41 47 80 58 45 60 44 72 77 ...
## $ HOSPITAL.COST: num 660293 809130 362231 629990 444876 ...
## - attr(*, "comment")= chr "glbObsNew"
## PTID BODY.WEIGHT HOSPITAL.COST
## 1 1 49 660293
## 2 2 41 809130
## 3 3 47 362231
## 4 4 80 629990
## 5 5 58 444876
## 6 6 45 372357
## PTID BODY.WEIGHT HOSPITAL.COST
## 81 81 15 119935.4
## 113 113 60 138093.0
## 157 157 13 132226.0
## 173 173 9 137273.0
## 237 237 52 209886.0
## 248 248 71 248112.0
## PTID BODY.WEIGHT HOSPITAL.COST
## 243 243 62 73682
## 244 244 69 295155
## 245 245 57 200321
## 246 246 58 191188
## 247 247 65 202807
## 248 248 71 248112
## 'data.frame': 248 obs. of 3 variables:
## $ PTID : int 1 2 3 4 5 6 7 8 9 10 ...
## $ BODY.WEIGHT : int 49 41 47 80 58 45 60 44 72 77 ...
## $ HOSPITAL.COST: num 660293 809130 362231 629990 444876 ...
## - attr(*, "comment")= chr "glbObsTrn"## Warning: glbObsTrn same as glbObsAll## Warning: glbObsNew same as glbObsAll## [1] "Partition stats:"## Loading required package: sqldf
## Loading required package: gsubfn
## Loading required package: proto
## Loading required package: RSQLite
## Loading required package: DBI
## Loading required package: tcltk## HOSPITAL.COST.cut.fctr .src .n
## 1 (4.53e+04,3.27e+05] Test 217
## 2 (4.53e+04,3.27e+05] Train 217
## 3 (3.27e+05,6.07e+05] Test 26
## 4 (3.27e+05,6.07e+05] Train 26
## 5 (6.07e+05,8.88e+05] Test 5
## 6 (6.07e+05,8.88e+05] Train 5
## HOSPITAL.COST.cut.fctr .src .n
## 1 (4.53e+04,3.27e+05] Test 217
## 2 (4.53e+04,3.27e+05] Train 217
## 3 (3.27e+05,6.07e+05] Test 26
## 4 (3.27e+05,6.07e+05] Train 26
## 5 (6.07e+05,8.88e+05] Test 5
## 6 (6.07e+05,8.88e+05] Train 5
## .src .n
## 1 Test 248
## 2 Train 248## Loading required package: lazyeval
## Loading required package: gdata
## gdata: read.xls support for 'XLS' (Excel 97-2004) files ENABLED.
##
## gdata: read.xls support for 'XLSX' (Excel 2007+) files ENABLED.
##
## Attaching package: 'gdata'
##
## The following objects are masked from 'package:dplyr':
##
## combine, first, last
##
## The following object is masked from 'package:stats':
##
## nobs
##
## The following object is masked from 'package:utils':
##
## object.size## [1] "Skipping duplicates check since glb_split_newdata_method == 'copy'"## label step_major step_minor label_minor bgn end elapsed
## 1 import.data 1 0 0 10.550 19.452 8.902
## 2 inspect.data 2 0 0 19.453 NA NAStep 2.0: inspect data
## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
## [1] "numeric data missing in glbObsAll: "
## named integer(0)
## [1] "numeric data w/ 0s in glbObsAll: "
## named integer(0)
## [1] "numeric data w/ Infs in glbObsAll: "
## named integer(0)
## [1] "numeric data w/ NaNs in glbObsAll: "
## named integer(0)
## [1] "string data missing in glbObsAll: "
## named list()## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
## label step_major step_minor label_minor bgn end elapsed
## 2 inspect.data 2 0 0 19.453 23.335 3.882
## 3 scrub.data 2 1 1 23.335 NA NAStep 2.1: scrub data
## [1] "numeric data missing in : "
## named integer(0)
## [1] "numeric data w/ 0s in : "
## named integer(0)
## [1] "numeric data w/ Infs in : "
## named integer(0)
## [1] "numeric data w/ NaNs in : "
## named integer(0)
## [1] "string data missing in : "
## named list()## label step_major step_minor label_minor bgn end elapsed
## 3 scrub.data 2 1 1 23.335 24.508 1.173
## 4 transform.data 2 2 2 24.509 NA NAStep 2.2: transform data
## [1] "Creating new feature: .pos..."
## [1] "Creating new feature: HospCost.cut.fctr..."## label step_major step_minor label_minor bgn end elapsed
## 4 transform.data 2 2 2 24.509 24.558 0.049
## 5 extract.features 3 0 0 24.558 NA NAStep 3.0: extract features
## label step_major step_minor label_minor bgn end
## 1 extract.features_bgn 1 0 0 24.615 NA
## elapsed
## 1 NA## label step_major step_minor label_minor
## 1 extract.features_bgn 1 0 0
## 2 extract.features_factorize.str.vars 2 0 0
## bgn end elapsed
## 1 24.615 24.625 0.011
## 2 24.626 NA NA## .src
## ".src"## label step_major step_minor label_minor
## 2 extract.features_factorize.str.vars 2 0 0
## 3 extract.features_end 3 0 0
## bgn end elapsed
## 2 24.626 24.648 0.022
## 3 24.649 NA NA## label step_major step_minor label_minor
## 2 extract.features_factorize.str.vars 2 0 0
## 1 extract.features_bgn 1 0 0
## bgn end elapsed duration
## 2 24.626 24.648 0.022 0.022
## 1 24.615 24.625 0.011 0.010
## [1] "Total Elapsed Time: 24.648 secs"
## time trans "bgn " "fit.data.training.all " "predict.data.new " "end "
## 0.0000 multiple enabled transitions: data.training.all data.new model.selected firing: data.training.all
## 1.0000 1 2 1 0 0
## 1.0000 multiple enabled transitions: data.training.all data.new model.selected model.final data.training.all.prediction firing: data.new
## 2.0000 2 1 1 1 0
## label step_major step_minor label_minor bgn end
## 5 extract.features 3 0 0 24.558 25.968
## 6 manage.missing.data 3 1 1 25.969 NA
## elapsed
## 5 1.41
## 6 NAStep 3.1: manage missing data
## [1] "numeric data missing in : "
## named integer(0)
## [1] "numeric data w/ 0s in : "
## named integer(0)
## [1] "numeric data w/ Infs in : "
## named integer(0)
## [1] "numeric data w/ NaNs in : "
## named integer(0)
## [1] "string data missing in : "
## named list()## [1] "numeric data missing in : "
## named integer(0)
## [1] "numeric data w/ 0s in : "
## named integer(0)
## [1] "numeric data w/ Infs in : "
## named integer(0)
## [1] "numeric data w/ NaNs in : "
## named integer(0)
## [1] "string data missing in : "
## named list()## label step_major step_minor label_minor bgn end
## 6 manage.missing.data 3 1 1 25.969 26.397
## 7 cluster.data 3 2 2 26.398 NA
## elapsed
## 6 0.429
## 7 NAStep 3.2: cluster data
## label step_major step_minor label_minor bgn end
## 7 cluster.data 3 2 2 26.398 26.44
## 8 partition.data.training 4 0 0 26.440 NA
## elapsed
## 7 0.042
## 8 NAStep 4.0: partition data training
## [1] "Newdata contains non-NA data for HOSPITAL.COST; setting OOB to Newdata"## HospCost.cut.fctr .n.Fit .n.OOB .n.Tst .freqRatio.Fit .freqRatio.OOB
## 1 (1e+05,2e+05] 151 151 151 0.6088710 0.6088710
## 2 (2e+05,3e+05] 44 44 44 0.1774194 0.1774194
## 3 (3e+05,9e+05] 34 34 34 0.1370968 0.1370968
## 4 [0,1e+05] 19 19 19 0.0766129 0.0766129
## .freqRatio.Tst
## 1 0.6088710
## 2 0.1774194
## 3 0.1370968
## 4 0.0766129## [1] "glbObsAll: "## [1] 496 8## [1] "glbObsTrn: "## [1] 248 8## [1] "glbObsFit: "## [1] 248 7## [1] "glbObsOOB: "## [1] 248 7## [1] "glbObsNew: "## [1] 248 7## Warning in rm(split): object 'split' not found## label step_major step_minor label_minor bgn end
## 8 partition.data.training 4 0 0 26.44 26.58
## 9 select.features 5 0 0 26.58 NA
## elapsed
## 8 0.14
## 9 NAStep 5.0: select features
## id cor.y exclude.as.feat cor.y.abs
## HospCost.cut.fctr HospCost.cut.fctr 0.8585462 1 0.8585462
## PTID PTID -0.4989370 1 0.4989370
## .pos .pos -0.4989370 0 0.4989370
## BODY.WEIGHT BODY.WEIGHT 0.3483585 0 0.3483585
## .rnorm .rnorm 0.0695725 0 0.0695725## Loading required package: reshape2## id cor.y exclude.as.feat cor.y.abs
## HospCost.cut.fctr HospCost.cut.fctr 0.8585462 1 0.8585462
## BODY.WEIGHT BODY.WEIGHT 0.3483585 0 0.3483585
## .rnorm .rnorm 0.0695725 0 0.0695725
## .pos .pos -0.4989370 0 0.4989370
## PTID PTID -0.4989370 1 0.4989370
## cor.high.X freqRatio percentUnique zeroVar nzv
## HospCost.cut.fctr NA 3.431818 1.612903 FALSE FALSE
## BODY.WEIGHT NA 1.444444 29.435484 FALSE FALSE
## .rnorm NA 1.000000 100.000000 FALSE FALSE
## .pos NA 1.000000 100.000000 FALSE FALSE
## PTID NA 1.000000 100.000000 FALSE FALSE
## is.cor.y.abs.low
## HospCost.cut.fctr FALSE
## BODY.WEIGHT FALSE
## .rnorm FALSE
## .pos FALSE
## PTID FALSE## Warning in myplot_scatter(plt_feats_df, "percentUnique", "freqRatio",
## colorcol_name = "nzv", : converting nzv to class:factor## Warning: Removed 4 rows containing missing values (geom_point).## Warning: Removed 4 rows containing missing values (geom_point).## Warning: Removed 4 rows containing missing values (geom_point).
## [1] id cor.y exclude.as.feat cor.y.abs
## [5] cor.high.X freqRatio percentUnique zeroVar
## [9] nzv is.cor.y.abs.low
## <0 rows> (or 0-length row.names)
## [1] "numeric data missing in : "
## named integer(0)
## [1] "numeric data w/ 0s in : "
## named integer(0)
## [1] "numeric data w/ Infs in : "
## named integer(0)
## [1] "numeric data w/ NaNs in : "
## named integer(0)
## [1] "string data missing in : "
## .lcn
## 0## [1] "glb_feats_df:"## [1] 5 12## id exclude.as.feat rsp_var
## HOSPITAL.COST HOSPITAL.COST TRUE TRUE## id cor.y exclude.as.feat cor.y.abs cor.high.X
## PTID PTID -0.498937 TRUE 0.498937 NA
## HOSPITAL.COST HOSPITAL.COST NA TRUE NA NA
## freqRatio percentUnique zeroVar nzv is.cor.y.abs.low
## PTID 1 100 FALSE FALSE FALSE
## HOSPITAL.COST NA NA NA NA NA
## interaction.feat shapiro.test.p.value rsp_var_raw id_var
## PTID NA NA FALSE TRUE
## HOSPITAL.COST NA NA NA NA
## rsp_var
## PTID NA
## HOSPITAL.COST TRUE## [1] "glb_feats_df vs. glbObsAll: "## character(0)## [1] "glbObsAll vs. glb_feats_df: "## character(0)## label step_major step_minor label_minor bgn end elapsed
## 9 select.features 5 0 0 26.58 28.339 1.759
## 10 fit.models 6 0 0 28.34 NA NAStep 6.0: fit models
fit.models_0_chunk_df <- myadd_chunk(NULL, "fit.models_0_bgn", label.minor = "setup")## label step_major step_minor label_minor bgn end elapsed
## 1 fit.models_0_bgn 1 0 setup 28.815 NA NA# load(paste0(glb_out_pfx, "dsk.RData"))
get_model_sel_frmla <- function() {
model_evl_terms <- c(NULL)
# min.aic.fit might not be avl
lclMdlEvlCriteria <-
glbMdlMetricsEval[glbMdlMetricsEval %in% names(glb_models_df)]
for (metric in lclMdlEvlCriteria)
model_evl_terms <- c(model_evl_terms,
ifelse(length(grep("max", metric)) > 0, "-", "+"), metric)
if (glb_is_classification && glb_is_binomial)
model_evl_terms <- c(model_evl_terms, "-", "opt.prob.threshold.OOB")
model_sel_frmla <- as.formula(paste(c("~ ", model_evl_terms), collapse = " "))
return(model_sel_frmla)
}
get_dsp_models_df <- function() {
dsp_models_cols <- c("id",
glbMdlMetricsEval[glbMdlMetricsEval %in% names(glb_models_df)],
grep("opt.", names(glb_models_df), fixed = TRUE, value = TRUE))
dsp_models_df <-
#orderBy(get_model_sel_frmla(), glb_models_df)[, c("id", glbMdlMetricsEval)]
orderBy(get_model_sel_frmla(), glb_models_df)[, dsp_models_cols]
nCvMdl <- sapply(glb_models_lst, function(mdl) nrow(mdl$results))
nParams <- sapply(glb_models_lst, function(mdl) ifelse(mdl$method == "custom", 0,
nrow(subset(modelLookup(mdl$method), parameter != "parameter"))))
# nCvMdl <- nCvMdl[names(nCvMdl) != "avNNet"]
# nParams <- nParams[names(nParams) != "avNNet"]
if (length(cvMdlProblems <- nCvMdl[nCvMdl <= nParams]) > 0) {
print("Cross Validation issues:")
warning("Cross Validation issues:")
print(cvMdlProblems)
}
pltMdls <- setdiff(names(nCvMdl), names(cvMdlProblems))
pltMdls <- setdiff(pltMdls, names(nParams[nParams == 0]))
# length(pltMdls) == 21
png(paste0(glb_out_pfx, "bestTune.png"), width = 480 * 2, height = 480 * 4)
grid.newpage()
pushViewport(viewport(layout = grid.layout(ceiling(length(pltMdls) / 2.0), 2)))
pltIx <- 1
for (mdlId in pltMdls) {
print(ggplot(glb_models_lst[[mdlId]], highBestTune = TRUE) + labs(title = mdlId),
vp = viewport(layout.pos.row = ceiling(pltIx / 2.0),
layout.pos.col = ((pltIx - 1) %% 2) + 1))
pltIx <- pltIx + 1
}
dev.off()
if (all(row.names(dsp_models_df) != dsp_models_df$id))
row.names(dsp_models_df) <- dsp_models_df$id
return(dsp_models_df)
}
#get_dsp_models_df()
if (glb_is_classification && glb_is_binomial &&
(length(unique(glbObsFit[, glb_rsp_var])) < 2))
stop("glbObsFit$", glb_rsp_var, ": contains less than 2 unique values: ",
paste0(unique(glbObsFit[, glb_rsp_var]), collapse=", "))
max_cor_y_x_vars <- orderBy(~ -cor.y.abs,
subset(glb_feats_df, (exclude.as.feat == 0) & !nzv & !is.cor.y.abs.low &
is.na(cor.high.X)))[1:2, "id"]
max_cor_y_x_vars <- max_cor_y_x_vars[!is.na(max_cor_y_x_vars)]
if (!is.null(glb_Baseline_mdl_var)) {
if ((max_cor_y_x_vars[1] != glb_Baseline_mdl_var) &
(glb_feats_df[glb_feats_df$id == max_cor_y_x_vars[1], "cor.y.abs"] >
glb_feats_df[glb_feats_df$id == glb_Baseline_mdl_var, "cor.y.abs"]))
stop(max_cor_y_x_vars[1], " has a higher correlation with ", glb_rsp_var,
" than the Baseline var: ", glb_Baseline_mdl_var)
}
glb_model_type <- ifelse(glb_is_regression, "regression", "classification")
# Model specs
c("id.prefix", "method", "type",
# trainControl params
"preProc.method", "cv.n.folds", "cv.n.repeats", "summary.fn",
# train params
"metric", "metric.maximize", "tune.df")## [1] "id.prefix" "method" "type"
## [4] "preProc.method" "cv.n.folds" "cv.n.repeats"
## [7] "summary.fn" "metric" "metric.maximize"
## [10] "tune.df"# Baseline
if (!is.null(glb_Baseline_mdl_var)) {
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Baseline"), major.inc = FALSE,
label.minor = "mybaseln_classfr")
ret_lst <- myfit_mdl(mdl_id="Baseline",
model_method="mybaseln_classfr",
indep_vars_vctr=glb_Baseline_mdl_var,
rsp_var=glb_rsp_var,
fit_df=glbObsFit, OOB_df=glbObsOOB)
}
# Most Frequent Outcome "MFO" model: mean(y) for regression
# Not using caret's nullModel since model stats not avl
# Cannot use rpart for multinomial classification since it predicts non-MFO
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "MFO"), major.inc = FALSE,
label.minor = "myMFO_classfr")## label step_major step_minor label_minor bgn end
## 1 fit.models_0_bgn 1 0 setup 28.815 28.845
## 2 fit.models_0_MFO 1 1 myMFO_classfr 28.846 NA
## elapsed
## 1 0.03
## 2 NAret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = "MFO", type = glb_model_type, trainControl.method = "none",
train.method = ifelse(glb_is_regression, "lm", "myMFO_classfr"))),
indep_vars = ".rnorm", rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)## [1] "fitting model: MFO###lm"
## [1] " indep_vars: .rnorm"
## Fitting parameter = none on full training set
##
## Call:
## lm(formula = .outcome ~ ., data = dat)
##
## Residuals:
## Min 1Q Median 3Q Max
## -148847 -68217 -33762 25863 692280
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 199137 7790 25.562 <2e-16 ***
## .rnorm 8430 7707 1.094 0.275
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 122500 on 246 degrees of freedom
## Multiple R-squared: 0.00484, Adjusted R-squared: 0.000795
## F-statistic: 1.197 on 1 and 246 DF, p-value: 0.2751
##
## id feats max.nTuningRuns min.elapsedtime.everything
## 1 MFO###lm .rnorm 0 0.553
## min.elapsedtime.final max.R.sq.fit min.RMSE.fit max.Adj.R.sq.fit
## 1 0.003 0.004840332 122043.6 0.0007949678
## max.R.sq.OOB min.RMSE.OOB max.Adj.R.sq.OOB
## 1 -0.007501072 122798.1 -0.0115966if (glb_is_classification) {
# "random" model - only for classification;
# none needed for regression since it is same as MFO
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Random"), major.inc = FALSE,
label.minor = "myrandom_classfr")
#stop(here"); glb2Sav(); all.equal(glb_models_df, sav_models_df)
ret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = "Random", type = glb_model_type, trainControl.method = "none",
train.method = "myrandom_classfr")),
indep_vars = ".rnorm", rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)
}
# Max.cor.Y
# Check impact of cv
# rpart is not a good candidate since caret does not optimize cp (only tuning parameter of rpart) well
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Max.cor.Y.rcv.*X*"), major.inc = FALSE,
label.minor = "glmnet")## label step_major step_minor label_minor
## 2 fit.models_0_MFO 1 1 myMFO_classfr
## 3 fit.models_0_Max.cor.Y.rcv.*X* 1 2 glmnet
## bgn end elapsed
## 2 28.846 30.508 1.662
## 3 30.509 NA NAret_lst <- myfit_mdl(mdl_specs_lst=myinit_mdl_specs_lst(mdl_specs_lst=list(
id.prefix="Max.cor.Y.rcv.1X1", type=glb_model_type, trainControl.method="none",
train.method="glmnet")),
indep_vars=max_cor_y_x_vars, rsp_var=glb_rsp_var,
fit_df=glbObsFit, OOB_df=glbObsOOB)## [1] "fitting model: Max.cor.Y.rcv.1X1###glmnet"
## [1] " indep_vars: .pos,BODY.WEIGHT"## Loading required package: glmnet
## Loading required package: Matrix
## Loaded glmnet 2.0-2## Fitting alpha = 0.1, lambda = 1221 on full training set
## Length Class Mode
## a0 72 -none- numeric
## beta 144 dgCMatrix S4
## df 72 -none- numeric
## dim 2 -none- numeric
## lambda 72 -none- numeric
## dev.ratio 72 -none- numeric
## nulldev 1 -none- numeric
## npasses 1 -none- numeric
## jerr 1 -none- numeric
## offset 1 -none- logical
## call 5 -none- call
## nobs 1 -none- numeric
## lambdaOpt 1 -none- numeric
## xNames 2 -none- character
## problemType 1 -none- character
## tuneValue 2 data.frame list
## obsLevels 1 -none- logical
## [1] "min lambda > lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243479.3476 -372.0095 1265.5046
## [1] "max lambda < lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243528.7233 -372.3666 1266.5484
## id feats max.nTuningRuns
## 1 Max.cor.Y.rcv.1X1###glmnet .pos,BODY.WEIGHT 0
## min.elapsedtime.everything min.elapsedtime.final max.R.sq.fit
## 1 0.924 0.01 0.3034957
## min.RMSE.fit max.Adj.R.sq.fit max.R.sq.OOB min.RMSE.OOB max.Adj.R.sq.OOB
## 1 102101.2 0.2978099 0.3034864 102101.8 0.2978006if (glbMdlCheckRcv) {
# rcv_n_folds == 1 & rcv_n_repeats > 1 crashes
for (rcv_n_folds in seq(3, glb_rcv_n_folds + 2, 2))
for (rcv_n_repeats in seq(1, glb_rcv_n_repeats + 2, 2)) {
# Experiment specific code to avoid caret crash
# lcl_tune_models_df <- rbind(data.frame()
# ,data.frame(method = "glmnet", parameter = "alpha",
# vals = "0.100 0.325 0.550 0.775 1.000")
# ,data.frame(method = "glmnet", parameter = "lambda",
# vals = "9.342e-02")
# )
ret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst =
list(
id.prefix = paste0("Max.cor.Y.rcv.", rcv_n_folds, "X", rcv_n_repeats),
type = glb_model_type,
# tune.df = lcl_tune_models_df,
trainControl.method = "repeatedcv",
trainControl.number = rcv_n_folds,
trainControl.repeats = rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.method = "glmnet", train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize)),
indep_vars = max_cor_y_x_vars, rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)
}
# Add parallel coordinates graph of glb_models_df[, glbMdlMetricsEval] to evaluate cv parameters
tmp_models_cols <- c("id", "max.nTuningRuns",
glbMdlMetricsEval[glbMdlMetricsEval %in% names(glb_models_df)],
grep("opt.", names(glb_models_df), fixed = TRUE, value = TRUE))
print(myplot_parcoord(obs_df = subset(glb_models_df,
grepl("Max.cor.Y.rcv.", id, fixed = TRUE),
select = -feats)[, tmp_models_cols],
id_var = "id"))
}
# Useful for stacking decisions
# fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
# paste0("fit.models_0_", "Max.cor.Y[rcv.1X1.cp.0|]"), major.inc = FALSE,
# label.minor = "rpart")
#
# ret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
# id.prefix = "Max.cor.Y.rcv.1X1.cp.0", type = glb_model_type, trainControl.method = "none",
# train.method = "rpart",
# tune.df=data.frame(method="rpart", parameter="cp", min=0.0, max=0.0, by=0.1))),
# indep_vars=max_cor_y_x_vars, rsp_var=glb_rsp_var,
# fit_df=glbObsFit, OOB_df=glbObsOOB)
#stop(here"); glb2Sav(); all.equal(glb_models_df, sav_models_df)
# if (glb_is_regression || glb_is_binomial) # For multinomials this model will be run next by default
ret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = "Max.cor.Y",
type = glb_model_type, trainControl.method = "repeatedcv",
trainControl.number = glb_rcv_n_folds,
trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
trainControl.allowParallel = glbMdlAllowParallel,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = "rpart")),
indep_vars = max_cor_y_x_vars, rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)## [1] "fitting model: Max.cor.Y##rcv#rpart"
## [1] " indep_vars: .pos,BODY.WEIGHT"## Loading required package: rpart## Warning in nominalTrainWorkflow(x = x, y = y, wts = weights, info =
## trainInfo, : There were missing values in resampled performance measures.## Aggregating results
## Selecting tuning parameters
## Fitting cp = 0.00918 on full training set## Loading required package: rpart.plot
## Call:
## rpart(formula = .outcome ~ ., control = list(minsplit = 20, minbucket = 7,
## cp = 0, maxcompete = 4, maxsurrogate = 5, usesurrogate = 2,
## surrogatestyle = 0, maxdepth = 30, xval = 0))
## n= 248
##
## CP nsplit rel error
## 1 0.579415409 0 1.0000000
## 2 0.089707294 1 0.4205846
## 3 0.018069000 2 0.3308773
## 4 0.009178547 3 0.3128083
##
## Variable importance
## .pos BODY.WEIGHT
## 96 4
##
## Node number 1: 248 observations, complexity param=0.5794154
## mean=198723.3, MSE=1.49671e+10
## left son=2 (211 obs) right son=3 (37 obs)
## Primary splits:
## .pos < 74 to the right, improve=0.5794154, (0 missing)
## BODY.WEIGHT < 38.5 to the left, improve=0.1579872, (0 missing)
##
## Node number 2: 211 observations, complexity param=0.018069
## mean=159727, MSE=3.111513e+09
## left son=4 (120 obs) right son=5 (91 obs)
## Primary splits:
## BODY.WEIGHT < 38.5 to the left, improve=0.10215730, (0 missing)
## .pos < 258 to the right, improve=0.03594479, (0 missing)
## Surrogate splits:
## .pos < 114 to the right, agree=0.645, adj=0.176, (0 split)
##
## Node number 3: 37 observations, complexity param=0.08970729
## mean=421107.7, MSE=2.444902e+10
## left son=6 (22 obs) right son=7 (15 obs)
## Primary splits:
## .pos < 30 to the right, improve=0.36808990, (0 missing)
## BODY.WEIGHT < 50.5 to the right, improve=0.05139237, (0 missing)
## Surrogate splits:
## BODY.WEIGHT < 50.5 to the right, agree=0.649, adj=0.133, (0 split)
##
## Node number 4: 120 observations
## mean=144201.3, MSE=2.268239e+09
##
## Node number 5: 91 observations
## mean=180200.4, MSE=3.486498e+09
##
## Node number 6: 22 observations
## mean=342775.2, MSE=8.049477e+09
##
## Node number 7: 15 observations
## mean=535995.3, MSE=2.630306e+10
##
## n= 248
##
## node), split, n, deviance, yval
## * denotes terminal node
##
## 1) root 248 3.711840e+12 198723.3
## 2) .pos>=74 211 6.565292e+11 159727.0
## 4) BODY.WEIGHT< 38.5 120 2.721887e+11 144201.3 *
## 5) BODY.WEIGHT>=38.5 91 3.172713e+11 180200.4 *
## 3) .pos< 74 37 9.046136e+11 421107.7
## 6) .pos>=30 22 1.770885e+11 342775.2 *
## 7) .pos< 30 15 3.945460e+11 535995.3 *
## id feats max.nTuningRuns
## 1 Max.cor.Y##rcv#rpart .pos,BODY.WEIGHT 5
## min.elapsedtime.everything min.elapsedtime.final max.R.sq.fit
## 1 1.367 0.012 0.6871917
## min.RMSE.fit max.Adj.R.sq.fit max.R.sq.OOB min.RMSE.OOB max.Adj.R.sq.OOB
## 1 74482.18 NA 0.6785293 69364.86 NA
## max.Rsquared.fit min.RMSESD.fit max.RsquaredSD.fit
## 1 0.6438406 11491.25 0.07346218if ((length(glbFeatsDateTime) > 0) &&
(sum(grepl(paste(names(glbFeatsDateTime), "\\.day\\.minutes\\.poly\\.", sep = ""),
names(glbObsAll))) > 0)) {
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Max.cor.Y.Time.Poly"), major.inc = FALSE,
label.minor = "glmnet")
indepVars <- c(max_cor_y_x_vars,
grep(paste(names(glbFeatsDateTime), "\\.day\\.minutes\\.poly\\.", sep = ""),
names(glbObsAll), value = TRUE))
indepVars <- myadjust_interaction_feats(indepVars)
ret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = "Max.cor.Y.Time.Poly",
type = glb_model_type, trainControl.method = "repeatedcv",
trainControl.number = glb_rcv_n_folds, trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = "glmnet")),
indep_vars = indepVars,
rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)
}
if ((length(glbFeatsDateTime) > 0) &&
(sum(grepl(paste(names(glbFeatsDateTime), "\\.last[[:digit:]]", sep = ""),
names(glbObsAll))) > 0)) {
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Max.cor.Y.Time.Lag"), major.inc = FALSE,
label.minor = "glmnet")
indepVars <- c(max_cor_y_x_vars,
grep(paste(names(glbFeatsDateTime), "\\.last[[:digit:]]", sep = ""),
names(glbObsAll), value = TRUE))
indepVars <- myadjust_interaction_feats(indepVars)
ret_lst <- myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = "Max.cor.Y.Time.Lag",
type = glb_model_type,
tune.df = glbMdlTuneParams,
trainControl.method = "repeatedcv",
trainControl.number = glb_rcv_n_folds, trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = "glmnet")),
indep_vars = indepVars,
rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)
}
# Interactions.High.cor.Y
if (length(int_feats <- setdiff(setdiff(unique(glb_feats_df$cor.high.X), NA),
subset(glb_feats_df, nzv)$id)) > 0) {
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Interact.High.cor.Y"), major.inc = FALSE,
label.minor = "glmnet")
ret_lst <- myfit_mdl(mdl_specs_lst=myinit_mdl_specs_lst(mdl_specs_lst=list(
id.prefix="Interact.High.cor.Y",
type=glb_model_type, trainControl.method="repeatedcv",
trainControl.number=glb_rcv_n_folds, trainControl.repeats=glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method="glmnet")),
indep_vars=c(max_cor_y_x_vars, paste(max_cor_y_x_vars[1], int_feats, sep=":")),
rsp_var=glb_rsp_var,
fit_df=glbObsFit, OOB_df=glbObsOOB)
}
# Low.cor.X
fit.models_0_chunk_df <- myadd_chunk(fit.models_0_chunk_df,
paste0("fit.models_0_", "Low.cor.X"), major.inc = FALSE,
label.minor = "glmnet")## label step_major step_minor label_minor bgn
## 3 fit.models_0_Max.cor.Y.rcv.*X* 1 2 glmnet 30.509
## 4 fit.models_0_Low.cor.X 1 3 glmnet 35.502
## end elapsed
## 3 35.501 4.992
## 4 NA NAindep_vars <- subset(glb_feats_df, is.na(cor.high.X) & !nzv &
(exclude.as.feat != 1))[, "id"]
indep_vars <- myadjust_interaction_feats(indep_vars)
ret_lst <- myfit_mdl(mdl_specs_lst=myinit_mdl_specs_lst(mdl_specs_lst=list(
id.prefix="Low.cor.X",
type=glb_model_type,
tune.df = glbMdlTuneParams,
trainControl.method="repeatedcv",
trainControl.number=glb_rcv_n_folds, trainControl.repeats=glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method="glmnet")),
indep_vars=indep_vars, rsp_var=glb_rsp_var,
fit_df=glbObsFit, OOB_df=glbObsOOB)## [1] "fitting model: Low.cor.X##rcv#glmnet"
## [1] " indep_vars: BODY.WEIGHT,.rnorm,.pos"
## Aggregating results
## Selecting tuning parameters
## Fitting alpha = 0.775, lambda = 5666 on full training set
## Length Class Mode
## a0 59 -none- numeric
## beta 177 dgCMatrix S4
## df 59 -none- numeric
## dim 2 -none- numeric
## lambda 59 -none- numeric
## dev.ratio 59 -none- numeric
## nulldev 1 -none- numeric
## npasses 1 -none- numeric
## jerr 1 -none- numeric
## offset 1 -none- logical
## call 5 -none- call
## nobs 1 -none- numeric
## lambdaOpt 1 -none- numeric
## xNames 3 -none- character
## problemType 1 -none- character
## tuneValue 2 data.frame list
## obsLevels 1 -none- logical
## [1] "min lambda > lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243053.5207 -347.2823 1113.4773
## [1] "max lambda < lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243140.3487 -349.8186 1127.9216
## id feats max.nTuningRuns
## 1 Low.cor.X##rcv#glmnet BODY.WEIGHT,.rnorm,.pos 25
## min.elapsedtime.everything min.elapsedtime.final max.R.sq.fit
## 1 1.681 0.005 0.3010613
## min.RMSE.fit max.Adj.R.sq.fit max.R.sq.OOB min.RMSE.OOB max.Adj.R.sq.OOB
## 1 102975.3 0.2924678 0.3010532 102280 0.2924596
## max.Rsquared.fit min.RMSESD.fit max.RsquaredSD.fit
## 1 0.2997312 12581.36 0.04504516fit.models_0_chunk_df <-
myadd_chunk(fit.models_0_chunk_df, "fit.models_0_end", major.inc = FALSE,
label.minor = "teardown")## label step_major step_minor label_minor bgn end
## 4 fit.models_0_Low.cor.X 1 3 glmnet 35.502 38.714
## 5 fit.models_0_end 1 4 teardown 38.715 NA
## elapsed
## 4 3.212
## 5 NArm(ret_lst)
glb_chunks_df <- myadd_chunk(glb_chunks_df, "fit.models", major.inc = FALSE)## label step_major step_minor label_minor bgn end elapsed
## 10 fit.models 6 0 0 28.340 38.723 10.383
## 11 fit.models 6 1 1 38.724 NA NAfit.models_1_chunk_df <- myadd_chunk(NULL, "fit.models_1_bgn", label.minor="setup")## label step_major step_minor label_minor bgn end elapsed
## 1 fit.models_1_bgn 1 0 setup 39.852 NA NA#stop(here"); glb2Sav(); all.equal(glb_models_df, sav_models_df)
topindep_var <- NULL; interact_vars <- NULL;
for (mdl_id_pfx in names(glbMdlFamilies)) {
fit.models_1_chunk_df <-
myadd_chunk(fit.models_1_chunk_df, paste0("fit.models_1_", mdl_id_pfx),
major.inc = FALSE, label.minor = "setup")
indep_vars <- NULL;
if (grepl("\\.Interact", mdl_id_pfx)) {
if (is.null(topindep_var) && is.null(interact_vars)) {
# select best glmnet model upto now
dsp_models_df <- orderBy(model_sel_frmla <- get_model_sel_frmla(),
glb_models_df)
dsp_models_df <- subset(dsp_models_df,
grepl(".glmnet", id, fixed = TRUE))
bst_mdl_id <- dsp_models_df$id[1]
mdl_id_pfx <-
paste(c(head(unlist(strsplit(bst_mdl_id, "[.]")), -1), "Interact"),
collapse=".")
# select important features
if (is.null(bst_featsimp_df <-
myget_feats_importance(glb_models_lst[[bst_mdl_id]]))) {
warning("Base model for RFE.Interact: ", bst_mdl_id,
" has no important features")
next
}
topindep_ix <- 1
while (is.null(topindep_var) && (topindep_ix <= nrow(bst_featsimp_df))) {
topindep_var <- row.names(bst_featsimp_df)[topindep_ix]
if (grepl(".fctr", topindep_var, fixed=TRUE))
topindep_var <-
paste0(unlist(strsplit(topindep_var, ".fctr"))[1], ".fctr")
if (topindep_var %in% names(glbFeatsInteractionOnly)) {
topindep_var <- NULL; topindep_ix <- topindep_ix + 1
} else break
}
# select features with importance > max(10, importance of .rnorm) & is not highest
# combine factor dummy features to just the factor feature
if (length(pos_rnorm <-
grep(".rnorm", row.names(bst_featsimp_df), fixed=TRUE)) > 0)
imp_rnorm <- bst_featsimp_df[pos_rnorm, 1] else
imp_rnorm <- NA
imp_cutoff <- max(10, imp_rnorm, na.rm=TRUE)
interact_vars <-
tail(row.names(subset(bst_featsimp_df,
imp > imp_cutoff)), -1)
if (length(interact_vars) > 0) {
interact_vars <-
myadjust_interaction_feats(myextract_actual_feats(interact_vars))
interact_vars <-
interact_vars[!grepl(topindep_var, interact_vars, fixed=TRUE)]
}
### bid0_sp only
# interact_vars <- c(
# "biddable", "D.ratio.sum.TfIdf.wrds.n", "D.TfIdf.sum.stem.stop.Ratio", "D.sum.TfIdf",
# "D.TfIdf.sum.post.stop", "D.TfIdf.sum.post.stem", "D.ratio.wrds.stop.n.wrds.n", "D.chrs.uppr.n.log",
# "D.chrs.n.log", "color.fctr"
# # , "condition.fctr", "prdl.my.descr.fctr"
# )
# interact_vars <- setdiff(interact_vars, c("startprice.dgt2.is9", "color.fctr"))
###
indep_vars <- myextract_actual_feats(row.names(bst_featsimp_df))
indep_vars <- setdiff(indep_vars, topindep_var)
if (length(interact_vars) > 0) {
indep_vars <-
setdiff(indep_vars, myextract_actual_feats(interact_vars))
indep_vars <- c(indep_vars,
paste(topindep_var, setdiff(interact_vars, topindep_var),
sep = "*"))
} else indep_vars <- union(indep_vars, topindep_var)
}
}
if (is.null(indep_vars))
indep_vars <- glb_mdl_feats_lst[[mdl_id_pfx]]
if (is.null(indep_vars) && grepl("RFE\\.", mdl_id_pfx))
indep_vars <- myextract_actual_feats(predictors(rfe_fit_results))
if (is.null(indep_vars))
indep_vars <- subset(glb_feats_df, !nzv & (exclude.as.feat != 1))[, "id"]
if ((length(indep_vars) == 1) && (grepl("^%<d-%", indep_vars))) {
indep_vars <-
eval(parse(text = str_trim(unlist(strsplit(indep_vars, "%<d-%"))[2])))
}
indep_vars <- myadjust_interaction_feats(indep_vars)
if (grepl("\\.Interact", mdl_id_pfx)) {
# if (method != tail(unlist(strsplit(bst_mdl_id, "[.]")), 1)) next
if (is.null(glbMdlFamilies[[mdl_id_pfx]])) {
if (!is.null(glbMdlFamilies[["Best.Interact"]]))
glbMdlFamilies[[mdl_id_pfx]] <-
glbMdlFamilies[["Best.Interact"]]
}
}
if (!is.null(glbObsFitOutliers[[mdl_id_pfx]])) {
fitobs_df <- glbObsFit[!(glbObsFit[, glb_id_var] %in%
glbObsFitOutliers[[mdl_id_pfx]]), ]
} else fitobs_df <- glbObsFit
if (is.null(glbMdlFamilies[[mdl_id_pfx]]))
mdl_methods <- glbMdlMethods else
mdl_methods <- glbMdlFamilies[[mdl_id_pfx]]
for (method in mdl_methods) {
if (method %in% c("rpart", "rf")) {
# rpart: fubar's the tree
# rf: skip the scenario w/ .rnorm for speed
indep_vars <- setdiff(indep_vars, c(".rnorm"))
#mdl_id <- paste0(mdl_id_pfx, ".no.rnorm")
}
fit.models_1_chunk_df <- myadd_chunk(fit.models_1_chunk_df,
paste0("fit.models_1_", mdl_id_pfx), major.inc = FALSE,
label.minor = method)
ret_lst <-
myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = mdl_id_pfx,
type = glb_model_type,
tune.df = glbMdlTuneParams,
trainControl.method = "repeatedcv",
trainControl.number = glb_rcv_n_folds,
trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = method)),
indep_vars = indep_vars, rsp_var = glb_rsp_var,
fit_df = fitobs_df, OOB_df = glbObsOOB)
}
}## label step_major step_minor label_minor bgn end
## 1 fit.models_1_bgn 1 0 setup 39.852 39.866
## 2 fit.models_1_All.X 1 1 setup 39.867 NA
## elapsed
## 1 0.014
## 2 NA
## label step_major step_minor label_minor bgn end
## 2 fit.models_1_All.X 1 1 setup 39.867 39.875
## 3 fit.models_1_All.X 1 2 glmnet 39.876 NA
## elapsed
## 2 0.008
## 3 NA
## [1] "fitting model: All.X##rcv#glmnet"
## [1] " indep_vars: BODY.WEIGHT,.rnorm,.pos"
## Aggregating results
## Selecting tuning parameters
## Fitting alpha = 0.775, lambda = 5666 on full training set
## Length Class Mode
## a0 59 -none- numeric
## beta 177 dgCMatrix S4
## df 59 -none- numeric
## dim 2 -none- numeric
## lambda 59 -none- numeric
## dev.ratio 59 -none- numeric
## nulldev 1 -none- numeric
## npasses 1 -none- numeric
## jerr 1 -none- numeric
## offset 1 -none- logical
## call 5 -none- call
## nobs 1 -none- numeric
## lambdaOpt 1 -none- numeric
## xNames 3 -none- character
## problemType 1 -none- character
## tuneValue 2 data.frame list
## obsLevels 1 -none- logical
## [1] "min lambda > lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243053.5207 -347.2823 1113.4773
## [1] "max lambda < lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243140.3487 -349.8186 1127.9216
## id feats max.nTuningRuns
## 1 All.X##rcv#glmnet BODY.WEIGHT,.rnorm,.pos 25
## min.elapsedtime.everything min.elapsedtime.final max.R.sq.fit
## 1 2.674 0.006 0.3010613
## min.RMSE.fit max.Adj.R.sq.fit max.R.sq.OOB min.RMSE.OOB max.Adj.R.sq.OOB
## 1 102975.3 0.2924678 0.3010532 102280 0.2924596
## max.Rsquared.fit min.RMSESD.fit max.RsquaredSD.fit
## 1 0.2997312 12581.36 0.04504516# Check if other preProcess methods improve model performance
fit.models_1_chunk_df <-
myadd_chunk(fit.models_1_chunk_df, "fit.models_1_preProc", major.inc = FALSE,
label.minor = "preProc")## label step_major step_minor label_minor bgn end
## 3 fit.models_1_All.X 1 2 glmnet 39.876 44.085
## 4 fit.models_1_preProc 1 3 preProc 44.086 NA
## elapsed
## 3 4.209
## 4 NAmdl_id <- orderBy(get_model_sel_frmla(), glb_models_df)[1, "id"]
indep_vars_vctr <- trim(unlist(strsplit(glb_models_df[glb_models_df$id == mdl_id,
"feats"], "[,]")))
method <- tail(unlist(strsplit(mdl_id, "[.]")), 1)
mdl_id_pfx <- paste0(head(unlist(strsplit(mdl_id, "[.]")), -1), collapse = ".")
if (!is.null(glbObsFitOutliers[[mdl_id_pfx]])) {
fitobs_df <- glbObsFit[!(glbObsFit[, glb_id_var] %in%
glbObsFitOutliers[[mdl_id_pfx]]), ]
} else fitobs_df <- glbObsFit
for (prePr in glb_preproc_methods) {
# The operations are applied in this order:
# Box-Cox/Yeo-Johnson transformation, centering, scaling, range, imputation, PCA, ICA then spatial sign.
ret_lst <- myfit_mdl(mdl_specs_lst=myinit_mdl_specs_lst(mdl_specs_lst=list(
id.prefix=mdl_id_pfx,
type=glb_model_type, tune.df=glbMdlTuneParams,
trainControl.method="repeatedcv",
trainControl.number=glb_rcv_n_folds,
trainControl.repeats=glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method=method, train.preProcess=prePr)),
indep_vars=indep_vars_vctr, rsp_var=glb_rsp_var,
fit_df=fitobs_df, OOB_df=glbObsOOB)
}
# If (All|RFE).X.glm is less accurate than Low.Cor.X.glm
# check NA coefficients & filter appropriate terms in indep_vars_vctr
# if (method == "glm") {
# orig_glm <- glb_models_lst[[paste0(mdl_id, ".", model_method)]]$finalModel
# orig_glm <- glb_models_lst[["All.X.glm"]]$finalModel; print(summary(orig_glm))
# orig_glm <- glb_models_lst[["RFE.X.glm"]]$finalModel; print(summary(orig_glm))
# require(car)
# vif_orig_glm <- vif(orig_glm); print(vif_orig_glm)
# # if vif errors out with "there are aliased coefficients in the model"
# alias_orig_glm <- alias(orig_glm); alias_complete_orig_glm <- (alias_orig_glm$Complete > 0); alias_complete_orig_glm <- alias_complete_orig_glm[rowSums(alias_complete_orig_glm) > 0, colSums(alias_complete_orig_glm) > 0]; print(alias_complete_orig_glm)
# print(vif_orig_glm[!is.na(vif_orig_glm) & (vif_orig_glm == Inf)])
# print(which.max(vif_orig_glm))
# print(sort(vif_orig_glm[vif_orig_glm >= 1.0e+03], decreasing=TRUE))
# glbObsFit[c(1143, 3637, 3953, 4105), c("UniqueID", "Popular", "H.P.quandary", "Headline")]
# glb_feats_df[glb_feats_df$id %in% grep("[HSA]\\.chrs.n.log", glb_feats_df$id, value=TRUE) | glb_feats_df$cor.high.X %in% grep("[HSA]\\.chrs.n.log", glb_feats_df$id, value=TRUE), ]
# all.equal(glbObsAll$S.chrs.uppr.n.log, glbObsAll$A.chrs.uppr.n.log)
# cor(glbObsAll$S.T.herald, glbObsAll$S.T.tribun)
# mydspObs(Abstract.contains="[Dd]iar", cols=("Abstract"), all=TRUE)
# subset(glb_feats_df, cor.y.abs <= glb_feats_df[glb_feats_df$id == ".rnorm", "cor.y.abs"])
# corxx_mtrx <- cor(data.matrix(glbObsAll[, setdiff(names(glbObsAll), myfind_chr_cols_df(glbObsAll))]), use="pairwise.complete.obs"); abs_corxx_mtrx <- abs(corxx_mtrx); diag(abs_corxx_mtrx) <- 0
# which.max(abs_corxx_mtrx["S.T.tribun", ])
# abs_corxx_mtrx["A.npnct08.log", "S.npnct08.log"]
# step_glm <- step(orig_glm)
# }
# Since caret does not optimize rpart well
# if (method == "rpart")
# ret_lst <- myfit_mdl(mdl_id=paste0(mdl_id_pfx, ".cp.0"), model_method=method,
# indep_vars_vctr=indep_vars_vctr,
# model_type=glb_model_type,
# rsp_var=glb_rsp_var,
# fit_df=glbObsFit, OOB_df=glbObsOOB,
# n_cv_folds=0, tune_models_df=data.frame(parameter="cp", min=0.0, max=0.0, by=0.1))
# User specified
# Ensure at least 2 vars in each regression; else varImp crashes
# sav_models_lst <- glb_models_lst; sav_models_df <- glb_models_df; sav_featsimp_df <- glb_featsimp_df; all.equal(sav_featsimp_df, glb_featsimp_df)
# glb_models_lst <- sav_models_lst; glb_models_df <- sav_models_df; glm_featsimp_df <- sav_featsimp_df
# easier to exclude features
# require(gdata) # needed for trim
# mdl_id <- "";
# indep_vars_vctr <- head(subset(glb_models_df, grepl("All\\.X\\.", mdl_id), select=feats)
# , 1)[, "feats"]
# indep_vars_vctr <- trim(unlist(strsplit(indep_vars_vctr, "[,]")))
# indep_vars_vctr <- setdiff(indep_vars_vctr, ".rnorm")
# easier to include features
#stop(here"); sav_models_df <- glb_models_df; glb_models_df <- sav_models_df
# !_sp
# mdl_id <- "csm"; indep_vars_vctr <- c(NULL
# ,"prdline.my.fctr", "prdline.my.fctr:.clusterid.fctr"
# ,"prdline.my.fctr*biddable"
# #,"prdline.my.fctr*startprice.log"
# #,"prdline.my.fctr*startprice.diff"
# ,"prdline.my.fctr*condition.fctr"
# ,"prdline.my.fctr*D.terms.post.stop.n"
# #,"prdline.my.fctr*D.terms.post.stem.n"
# ,"prdline.my.fctr*cellular.fctr"
# # ,"<feat1>:<feat2>"
# )
# for (method in glbMdlMethods) {
# ret_lst <- myfit_mdl(mdl_id=mdl_id, model_method=method,
# indep_vars_vctr=indep_vars_vctr,
# model_type=glb_model_type,
# rsp_var=glb_rsp_var,
# fit_df=glbObsFit, OOB_df=glbObsOOB,
# n_cv_folds=glb_rcv_n_folds, tune_models_df=glbMdlTuneParams)
# csm_mdl_id <- paste0(mdl_id, ".", method)
# csm_featsimp_df <- myget_feats_importance(glb_models_lst[[paste0(mdl_id, ".",
# method)]]); print(head(csm_featsimp_df))
# }
###
# Ntv.1.lm <- lm(reformulate(indep_vars_vctr, glb_rsp_var), glbObsTrn); print(summary(Ntv.1.lm))
#glb_models_df[, "max.Accuracy.OOB", FALSE]
#varImp(glb_models_lst[["Low.cor.X.glm"]])
#orderBy(~ -Overall, varImp(glb_models_lst[["All.X.2.glm"]])$imp)
#orderBy(~ -Overall, varImp(glb_models_lst[["All.X.3.glm"]])$imp)
#glb_feats_df[grepl("npnct28", glb_feats_df$id), ]
# User specified bivariate models
# indep_vars_vctr_lst <- list()
# for (feat in setdiff(names(glbObsFit),
# union(glb_rsp_var, glbFeatsExclude)))
# indep_vars_vctr_lst[["feat"]] <- feat
# User specified combinatorial models
# indep_vars_vctr_lst <- list()
# combn_mtrx <- combn(c("<feat1_name>", "<feat2_name>", "<featn_name>"),
# <num_feats_to_choose>)
# for (combn_ix in 1:ncol(combn_mtrx))
# #print(combn_mtrx[, combn_ix])
# indep_vars_vctr_lst[[combn_ix]] <- combn_mtrx[, combn_ix]
# template for myfit_mdl
# rf is hard-coded in caret to recognize only Accuracy / Kappa evaluation metrics
# only for OOB in trainControl ?
# ret_lst <- myfit_mdl_fn(mdl_id=paste0(mdl_id_pfx, ""), model_method=method,
# indep_vars_vctr=indep_vars_vctr,
# rsp_var=glb_rsp_var,
# fit_df=glbObsFit, OOB_df=glbObsOOB,
# n_cv_folds=glb_rcv_n_folds, tune_models_df=glbMdlTuneParams,
# model_loss_mtrx=glbMdlMetric_terms,
# model_summaryFunction=glbMdlMetricSummaryFn,
# model_metric=glbMdlMetricSummary,
# model_metric_maximize=glbMdlMetricMaximize)
# Simplify a model
# fit_df <- glbObsFit; glb_mdl <- step(<complex>_mdl)
# Non-caret models
# rpart_area_mdl <- rpart(reformulate("Area", response=glb_rsp_var),
# data=glbObsFit, #method="class",
# control=rpart.control(cp=0.12),
# parms=list(loss=glbMdlMetric_terms))
# print("rpart_sel_wlm_mdl"); prp(rpart_sel_wlm_mdl)
#
print(glb_models_df)## id
## MFO###lm MFO###lm
## Max.cor.Y.rcv.1X1###glmnet Max.cor.Y.rcv.1X1###glmnet
## Max.cor.Y##rcv#rpart Max.cor.Y##rcv#rpart
## Low.cor.X##rcv#glmnet Low.cor.X##rcv#glmnet
## All.X##rcv#glmnet All.X##rcv#glmnet
## feats max.nTuningRuns
## MFO###lm .rnorm 0
## Max.cor.Y.rcv.1X1###glmnet .pos,BODY.WEIGHT 0
## Max.cor.Y##rcv#rpart .pos,BODY.WEIGHT 5
## Low.cor.X##rcv#glmnet BODY.WEIGHT,.rnorm,.pos 25
## All.X##rcv#glmnet BODY.WEIGHT,.rnorm,.pos 25
## min.elapsedtime.everything
## MFO###lm 0.553
## Max.cor.Y.rcv.1X1###glmnet 0.924
## Max.cor.Y##rcv#rpart 1.367
## Low.cor.X##rcv#glmnet 1.681
## All.X##rcv#glmnet 2.674
## min.elapsedtime.final max.R.sq.fit min.RMSE.fit
## MFO###lm 0.003 0.004840332 122043.65
## Max.cor.Y.rcv.1X1###glmnet 0.010 0.303495657 102101.17
## Max.cor.Y##rcv#rpart 0.012 0.687191703 74482.18
## Low.cor.X##rcv#glmnet 0.005 0.301061275 102975.27
## All.X##rcv#glmnet 0.006 0.301061275 102975.27
## max.Adj.R.sq.fit max.R.sq.OOB min.RMSE.OOB
## MFO###lm 0.0007949678 -0.007501072 122798.07
## Max.cor.Y.rcv.1X1###glmnet 0.2978099076 0.303486397 102101.85
## Max.cor.Y##rcv#rpart NA 0.678529263 69364.86
## Low.cor.X##rcv#glmnet 0.2924677657 0.301053181 102280.03
## All.X##rcv#glmnet 0.2924677657 0.301053181 102280.03
## max.Adj.R.sq.OOB max.Rsquared.fit
## MFO###lm -0.0115966 NA
## Max.cor.Y.rcv.1X1###glmnet 0.2978006 NA
## Max.cor.Y##rcv#rpart NA 0.6438406
## Low.cor.X##rcv#glmnet 0.2924596 0.2997312
## All.X##rcv#glmnet 0.2924596 0.2997312
## min.RMSESD.fit max.RsquaredSD.fit
## MFO###lm NA NA
## Max.cor.Y.rcv.1X1###glmnet NA NA
## Max.cor.Y##rcv#rpart 11491.25 0.07346218
## Low.cor.X##rcv#glmnet 12581.36 0.04504516
## All.X##rcv#glmnet 12581.36 0.04504516rm(ret_lst)
fit.models_1_chunk_df <-
myadd_chunk(fit.models_1_chunk_df, "fit.models_1_end", major.inc = FALSE,
label.minor = "teardown")## label step_major step_minor label_minor bgn end
## 4 fit.models_1_preProc 1 3 preProc 44.086 44.137
## 5 fit.models_1_end 1 4 teardown 44.137 NA
## elapsed
## 4 0.051
## 5 NAglb_chunks_df <- myadd_chunk(glb_chunks_df, "fit.models", major.inc = FALSE)## label step_major step_minor label_minor bgn end elapsed
## 11 fit.models 6 1 1 38.724 44.145 5.421
## 12 fit.models 6 2 2 44.146 NA NAfit.models_2_chunk_df <-
myadd_chunk(NULL, "fit.models_2_bgn", label.minor = "setup")## label step_major step_minor label_minor bgn end elapsed
## 1 fit.models_2_bgn 1 0 setup 44.88 NA NAplt_models_df <- glb_models_df[, -grep("SD|Upper|Lower", names(glb_models_df))]
for (var in grep("^min.", names(plt_models_df), value=TRUE)) {
plt_models_df[, sub("min.", "inv.", var)] <-
#ifelse(all(is.na(tmp <- plt_models_df[, var])), NA, 1.0 / tmp)
1.0 / plt_models_df[, var]
plt_models_df <- plt_models_df[ , -grep(var, names(plt_models_df))]
}
print(plt_models_df)## id
## MFO###lm MFO###lm
## Max.cor.Y.rcv.1X1###glmnet Max.cor.Y.rcv.1X1###glmnet
## Max.cor.Y##rcv#rpart Max.cor.Y##rcv#rpart
## Low.cor.X##rcv#glmnet Low.cor.X##rcv#glmnet
## All.X##rcv#glmnet All.X##rcv#glmnet
## feats max.nTuningRuns
## MFO###lm .rnorm 0
## Max.cor.Y.rcv.1X1###glmnet .pos,BODY.WEIGHT 0
## Max.cor.Y##rcv#rpart .pos,BODY.WEIGHT 5
## Low.cor.X##rcv#glmnet BODY.WEIGHT,.rnorm,.pos 25
## All.X##rcv#glmnet BODY.WEIGHT,.rnorm,.pos 25
## max.R.sq.fit max.Adj.R.sq.fit max.R.sq.OOB
## MFO###lm 0.004840332 0.0007949678 -0.007501072
## Max.cor.Y.rcv.1X1###glmnet 0.303495657 0.2978099076 0.303486397
## Max.cor.Y##rcv#rpart 0.687191703 NA 0.678529263
## Low.cor.X##rcv#glmnet 0.301061275 0.2924677657 0.301053181
## All.X##rcv#glmnet 0.301061275 0.2924677657 0.301053181
## max.Adj.R.sq.OOB max.Rsquared.fit
## MFO###lm -0.0115966 NA
## Max.cor.Y.rcv.1X1###glmnet 0.2978006 NA
## Max.cor.Y##rcv#rpart NA 0.6438406
## Low.cor.X##rcv#glmnet 0.2924596 0.2997312
## All.X##rcv#glmnet 0.2924596 0.2997312
## inv.elapsedtime.everything
## MFO###lm 1.8083183
## Max.cor.Y.rcv.1X1###glmnet 1.0822511
## Max.cor.Y##rcv#rpart 0.7315289
## Low.cor.X##rcv#glmnet 0.5948840
## All.X##rcv#glmnet 0.3739716
## inv.elapsedtime.final inv.RMSE.fit inv.RMSE.OOB
## MFO###lm 333.33333 8.193790e-06 8.143450e-06
## Max.cor.Y.rcv.1X1###glmnet 100.00000 9.794207e-06 9.794142e-06
## Max.cor.Y##rcv#rpart 83.33333 1.342603e-05 1.441652e-05
## Low.cor.X##rcv#glmnet 200.00000 9.711069e-06 9.777079e-06
## All.X##rcv#glmnet 166.66667 9.711069e-06 9.777079e-06print(myplot_radar(radar_inp_df=plt_models_df))## Warning: Removed 4 rows containing missing values (geom_path).## Warning: Removed 4 rows containing missing values (geom_point).## Warning: Removed 4 rows containing missing values (geom_text).
# print(myplot_radar(radar_inp_df=subset(plt_models_df,
# !(mdl_id %in% grep("random|MFO", plt_models_df$id, value=TRUE)))))
# Compute CI for <metric>SD
glb_models_df <- mutate(glb_models_df,
max.df = ifelse(max.nTuningRuns > 1, max.nTuningRuns - 1, NA),
min.sd2ci.scaler = ifelse(is.na(max.df), NA, qt(0.975, max.df)))
for (var in grep("SD", names(glb_models_df), value=TRUE)) {
# Does CI alredy exist ?
var_components <- unlist(strsplit(var, "SD"))
varActul <- paste0(var_components[1], var_components[2])
varUpper <- paste0(var_components[1], "Upper", var_components[2])
varLower <- paste0(var_components[1], "Lower", var_components[2])
if (varUpper %in% names(glb_models_df)) {
warning(varUpper, " already exists in glb_models_df")
# Assuming Lower also exists
next
}
print(sprintf("var:%s", var))
# CI is dependent on sample size in t distribution; df=n-1
glb_models_df[, varUpper] <- glb_models_df[, varActul] +
glb_models_df[, "min.sd2ci.scaler"] * glb_models_df[, var]
glb_models_df[, varLower] <- glb_models_df[, varActul] -
glb_models_df[, "min.sd2ci.scaler"] * glb_models_df[, var]
}## [1] "var:min.RMSESD.fit"
## [1] "var:max.RsquaredSD.fit"# Plot metrics with CI
plt_models_df <- glb_models_df[, "id", FALSE]
pltCI_models_df <- glb_models_df[, "id", FALSE]
for (var in grep("Upper", names(glb_models_df), value=TRUE)) {
var_components <- unlist(strsplit(var, "Upper"))
col_name <- unlist(paste(var_components, collapse=""))
plt_models_df[, col_name] <- glb_models_df[, col_name]
for (name in paste0(var_components[1], c("Upper", "Lower"), var_components[2]))
pltCI_models_df[, name] <- glb_models_df[, name]
}
build_statsCI_data <- function(plt_models_df) {
mltd_models_df <- melt(plt_models_df, id.vars="id")
mltd_models_df$data <- sapply(1:nrow(mltd_models_df),
function(row_ix) tail(unlist(strsplit(as.character(
mltd_models_df[row_ix, "variable"]), "[.]")), 1))
mltd_models_df$label <- sapply(1:nrow(mltd_models_df),
function(row_ix) head(unlist(strsplit(as.character(
mltd_models_df[row_ix, "variable"]),
paste0(".", mltd_models_df[row_ix, "data"]))), 1))
#print(mltd_models_df)
return(mltd_models_df)
}
mltd_models_df <- build_statsCI_data(plt_models_df)
mltdCI_models_df <- melt(pltCI_models_df, id.vars="id")
for (row_ix in 1:nrow(mltdCI_models_df)) {
for (type in c("Upper", "Lower")) {
if (length(var_components <- unlist(strsplit(
as.character(mltdCI_models_df[row_ix, "variable"]), type))) > 1) {
#print(sprintf("row_ix:%d; type:%s; ", row_ix, type))
mltdCI_models_df[row_ix, "label"] <- var_components[1]
mltdCI_models_df[row_ix, "data"] <-
unlist(strsplit(var_components[2], "[.]"))[2]
mltdCI_models_df[row_ix, "type"] <- type
break
}
}
}
wideCI_models_df <- reshape(subset(mltdCI_models_df, select=-variable),
timevar="type",
idvar=setdiff(names(mltdCI_models_df), c("type", "value", "variable")),
direction="wide")
#print(wideCI_models_df)
mrgdCI_models_df <- merge(wideCI_models_df, mltd_models_df, all.x=TRUE)
#print(mrgdCI_models_df)
# Merge stats back in if CIs don't exist
goback_vars <- c()
for (var in unique(mltd_models_df$label)) {
for (type in unique(mltd_models_df$data)) {
var_type <- paste0(var, ".", type)
# if this data is already present, next
if (var_type %in% unique(paste(mltd_models_df$label, mltd_models_df$data,
sep=".")))
next
#print(sprintf("var_type:%s", var_type))
goback_vars <- c(goback_vars, var_type)
}
}
if (length(goback_vars) > 0) {
mltd_goback_df <- build_statsCI_data(glb_models_df[, c("id", goback_vars)])
mltd_models_df <- rbind(mltd_models_df, mltd_goback_df)
}
# mltd_models_df <- merge(mltd_models_df, glb_models_df[, c("id", "model_method")],
# all.x=TRUE)
png(paste0(glb_out_pfx, "models_bar.png"), width=480*3, height=480*2)
#print(gp <- myplot_bar(mltd_models_df, "id", "value", colorcol_name="model_method") +
print(gp <- myplot_bar(df=mltd_models_df, xcol_name="id", ycol_names="value") +
geom_errorbar(data=mrgdCI_models_df,
mapping=aes(x=mdl_id, ymax=value.Upper, ymin=value.Lower), width=0.5) +
facet_grid(label ~ data, scales="free") +
theme(axis.text.x = element_text(angle = 90,vjust = 0.5)))## Warning: Removed 2 rows containing missing values (position_stack).## Warning: Removed 4 rows containing missing values (geom_errorbar).dev.off()## quartz_off_screen
## 2print(gp)## Warning: Removed 2 rows containing missing values (position_stack).## Warning: Removed 4 rows containing missing values (geom_errorbar).
dsp_models_cols <- c("id",
glbMdlMetricsEval[glbMdlMetricsEval %in% names(glb_models_df)],
grep("opt.", names(glb_models_df), fixed = TRUE, value = TRUE))
# if (glb_is_classification && glb_is_binomial)
# dsp_models_cols <- c(dsp_models_cols, "opt.prob.threshold.OOB")
print(dsp_models_df <- orderBy(get_model_sel_frmla(), glb_models_df)[, dsp_models_cols])## id min.RMSE.OOB max.R.sq.OOB max.Adj.R.sq.fit
## 3 Max.cor.Y##rcv#rpart 69364.86 0.678529263 NA
## 2 Max.cor.Y.rcv.1X1###glmnet 102101.85 0.303486397 0.2978099076
## 4 Low.cor.X##rcv#glmnet 102280.03 0.301053181 0.2924677657
## 5 All.X##rcv#glmnet 102280.03 0.301053181 0.2924677657
## 1 MFO###lm 122798.07 -0.007501072 0.0007949678
## min.RMSE.fit
## 3 74482.18
## 2 102101.17
## 4 102975.27
## 5 102975.27
## 1 122043.65print(myplot_radar(radar_inp_df = dsp_models_df))## Warning: Removed 1 rows containing missing values (geom_path).## Warning: Removed 1 rows containing missing values (geom_point).## Warning: Removed 1 rows containing missing values (geom_text).
print("Metrics used for model selection:"); print(get_model_sel_frmla())## [1] "Metrics used for model selection:"## ~+min.RMSE.OOB - max.R.sq.OOB - max.Adj.R.sq.fit + min.RMSE.fit
## <environment: 0x7faae517ea78>print(sprintf("Best model id: %s", dsp_models_df[1, "id"]))## [1] "Best model id: Max.cor.Y##rcv#rpart"glb_get_predictions <- function(df, mdl_id, rsp_var, prob_threshold_def=NULL, verbose=FALSE) {
mdl <- glb_models_lst[[mdl_id]]
clmnNames <- mygetPredictIds(rsp_var, mdl_id)
predct_var_name <- clmnNames$value
predct_prob_var_name <- clmnNames$prob
predct_accurate_var_name <- clmnNames$is.acc
predct_error_var_name <- clmnNames$err
predct_erabs_var_name <- clmnNames$err.abs
if (glb_is_regression) {
df[, predct_var_name] <- predict(mdl, newdata=df, type="raw")
if (verbose) print(myplot_scatter(df, glb_rsp_var, predct_var_name) +
facet_wrap(reformulate(glbFeatsCategory), scales = "free") +
stat_smooth(method="glm"))
df[, predct_error_var_name] <- df[, predct_var_name] - df[, glb_rsp_var]
if (verbose) print(myplot_scatter(df, predct_var_name, predct_error_var_name) +
#facet_wrap(reformulate(glbFeatsCategory), scales = "free") +
stat_smooth(method="auto"))
if (verbose) print(myplot_scatter(df, glb_rsp_var, predct_error_var_name) +
#facet_wrap(reformulate(glbFeatsCategory), scales = "free") +
stat_smooth(method="glm"))
df[, predct_erabs_var_name] <- abs(df[, predct_error_var_name])
if (verbose) print(head(orderBy(reformulate(c("-", predct_erabs_var_name)), df)))
df[, predct_accurate_var_name] <- (df[, glb_rsp_var] == df[, predct_var_name])
}
if (glb_is_classification && glb_is_binomial) {
prob_threshold <- glb_models_df[glb_models_df$id == mdl_id,
"opt.prob.threshold.OOB"]
if (is.null(prob_threshold) || is.na(prob_threshold)) {
warning("Using default probability threshold: ", prob_threshold_def)
if (is.null(prob_threshold <- prob_threshold_def))
stop("Default probability threshold is NULL")
}
df[, predct_prob_var_name] <- predict(mdl, newdata = df, type = "prob")[, 2]
df[, predct_var_name] <-
factor(levels(df[, glb_rsp_var])[
(df[, predct_prob_var_name] >=
prob_threshold) * 1 + 1], levels(df[, glb_rsp_var]))
# if (verbose) print(myplot_scatter(df, glb_rsp_var, predct_var_name) +
# facet_wrap(reformulate(glbFeatsCategory), scales = "free") +
# stat_smooth(method="glm"))
df[, predct_error_var_name] <- df[, predct_var_name] != df[, glb_rsp_var]
# if (verbose) print(myplot_scatter(df, predct_var_name, predct_error_var_name) +
# #facet_wrap(reformulate(glbFeatsCategory), scales = "free") +
# stat_smooth(method="auto"))
# if (verbose) print(myplot_scatter(df, glb_rsp_var, predct_error_var_name) +
# #facet_wrap(reformulate(glbFeatsCategory), scales = "free") +
# stat_smooth(method="glm"))
# if prediction is a TP (true +ve), measure distance from 1.0
tp <- which((df[, predct_var_name] == df[, glb_rsp_var]) &
(df[, predct_var_name] == levels(df[, glb_rsp_var])[2]))
df[tp, predct_erabs_var_name] <- abs(1 - df[tp, predct_prob_var_name])
#rowIx <- which.max(df[tp, predct_erabs_var_name]); df[tp, c(glb_id_var, glb_rsp_var, predct_var_name, predct_prob_var_name, predct_erabs_var_name)][rowIx, ]
# if prediction is a TN (true -ve), measure distance from 0.0
tn <- which((df[, predct_var_name] == df[, glb_rsp_var]) &
(df[, predct_var_name] == levels(df[, glb_rsp_var])[1]))
df[tn, predct_erabs_var_name] <- abs(0 - df[tn, predct_prob_var_name])
#rowIx <- which.max(df[tn, predct_erabs_var_name]); df[tn, c(glb_id_var, glb_rsp_var, predct_var_name, predct_prob_var_name, predct_erabs_var_name)][rowIx, ]
# if prediction is a FP (flse +ve), measure distance from 0.0
fp <- which((df[, predct_var_name] != df[, glb_rsp_var]) &
(df[, predct_var_name] == levels(df[, glb_rsp_var])[2]))
df[fp, predct_erabs_var_name] <- abs(0 - df[fp, predct_prob_var_name])
#rowIx <- which.max(df[fp, predct_erabs_var_name]); df[fp, c(glb_id_var, glb_rsp_var, predct_var_name, predct_prob_var_name, predct_erabs_var_name)][rowIx, ]
# if prediction is a FN (flse -ve), measure distance from 1.0
fn <- which((df[, predct_var_name] != df[, glb_rsp_var]) &
(df[, predct_var_name] == levels(df[, glb_rsp_var])[1]))
df[fn, predct_erabs_var_name] <- abs(1 - df[fn, predct_prob_var_name])
#rowIx <- which.max(df[fn, predct_erabs_var_name]); df[fn, c(glb_id_var, glb_rsp_var, predct_var_name, predct_prob_var_name, predct_erabs_var_name)][rowIx, ]
if (verbose) print(head(orderBy(reformulate(c("-", predct_erabs_var_name)), df)))
df[, predct_accurate_var_name] <- (df[, glb_rsp_var] == df[, predct_var_name])
}
if (glb_is_classification && !glb_is_binomial) {
df[, predct_var_name] <- predict(mdl, newdata = df, type = "raw")
df[, paste0(predct_var_name, ".prob")] <-
predict(mdl, newdata = df, type = "prob")
stop("Multinomial prediction error calculation needs to be implemented...")
}
return(df)
}
#stop(here"); glb2Sav(); glbObsAll <- savObsAll; glbObsTrn <- savObsTrn; glbObsFit <- savObsFit; glbObsOOB <- savObsOOB; sav_models_df <- glb_models_df; glb_models_df <- sav_models_df; glb_featsimp_df <- sav_featsimp_df
myget_category_stats <- function(obs_df, mdl_id, label) {
require(dplyr)
require(lazyeval)
predct_var_name <- mygetPredictIds(glb_rsp_var, mdl_id)$value
predct_error_var_name <- mygetPredictIds(glb_rsp_var, mdl_id)$err.abs
if (!predct_var_name %in% names(obs_df))
obs_df <- glb_get_predictions(obs_df, mdl_id, glb_rsp_var)
tmp_obs_df <- obs_df[, c(glbFeatsCategory, glb_rsp_var,
predct_var_name, predct_error_var_name)]
# tmp_obs_df <- obs_df %>%
# dplyr::select_(glbFeatsCategory, glb_rsp_var, predct_var_name, predct_error_var_name)
#dplyr::rename(startprice.log10.predict.RFE.X.glmnet.err=error_abs_OOB)
names(tmp_obs_df)[length(names(tmp_obs_df))] <- paste0("err.abs.", label)
ret_ctgry_df <- tmp_obs_df %>%
dplyr::group_by_(glbFeatsCategory) %>%
dplyr::summarise_(#interp(~sum(abs(var)), var=as.name(glb_rsp_var)),
interp(~sum(var), var=as.name(paste0("err.abs.", label))),
interp(~mean(var), var=as.name(paste0("err.abs.", label))),
interp(~n()))
names(ret_ctgry_df) <- c(glbFeatsCategory,
#paste0(glb_rsp_var, ".abs.", label, ".sum"),
paste0("err.abs.", label, ".sum"),
paste0("err.abs.", label, ".mean"),
paste0(".n.", label))
ret_ctgry_df <- dplyr::ungroup(ret_ctgry_df)
#colSums(ret_ctgry_df[, -grep(glbFeatsCategory, names(ret_ctgry_df))])
return(ret_ctgry_df)
}
#print(colSums((ctgry_df <- myget_category_stats(obs_df=glbObsFit, mdl_id="", label="fit"))[, -grep(glbFeatsCategory, names(ctgry_df))]))
if (!is.null(glb_mdl_ensemble)) {
fit.models_2_chunk_df <- myadd_chunk(fit.models_2_chunk_df,
paste0("fit.models_2_", mdl_id_pfx), major.inc = TRUE,
label.minor = "ensemble")
mdl_id_pfx <- "Ensemble"
if (#(glb_is_regression) |
((glb_is_classification) & (!glb_is_binomial)))
stop("Ensemble models not implemented yet for multinomial classification")
mygetEnsembleAutoMdlIds <- function() {
tmp_models_df <- orderBy(get_model_sel_frmla(), glb_models_df)
row.names(tmp_models_df) <- tmp_models_df$id
mdl_threshold_pos <-
min(which(grepl("MFO|Random|Baseline", tmp_models_df$id))) - 1
mdlIds <- tmp_models_df$id[1:mdl_threshold_pos]
return(mdlIds[!grepl("Ensemble", mdlIds)])
}
if (glb_mdl_ensemble == "auto") {
glb_mdl_ensemble <- mygetEnsembleAutoMdlIds()
mdl_id_pfx <- paste0(mdl_id_pfx, ".auto")
} else if (grepl("^%<d-%", glb_mdl_ensemble)) {
glb_mdl_ensemble <- eval(parse(text =
str_trim(unlist(strsplit(glb_mdl_ensemble, "%<d-%"))[2])))
}
for (mdl_id in glb_mdl_ensemble) {
if (!(mdl_id %in% names(glb_models_lst))) {
warning("Model ", mdl_id, " in glb_model_ensemble not found !")
next
}
glbObsFit <- glb_get_predictions(df = glbObsFit, mdl_id, glb_rsp_var)
glbObsOOB <- glb_get_predictions(df = glbObsOOB, mdl_id, glb_rsp_var)
}
#mdl_id_pfx <- "Ensemble.RFE"; mdlId <- paste0(mdl_id_pfx, ".glmnet")
#glb_mdl_ensemble <- gsub(mygetPredictIds$value, "", grep("RFE\\.X\\.(?!Interact)", row.names(glb_featsimp_df), perl = TRUE, value = TRUE), fixed = TRUE)
#varImp(glb_models_lst[[mdlId]])
#cor_df <- data.frame(cor=cor(glbObsFit[, glb_rsp_var], glbObsFit[, paste(mygetPredictIds$value, glb_mdl_ensemble)], use="pairwise.complete.obs"))
#glbObsFit <- glb_get_predictions(df=glbObsFit, "Ensemble.glmnet", glb_rsp_var);print(colSums((ctgry_df <- myget_category_stats(obs_df=glbObsFit, mdl_id="Ensemble.glmnet", label="fit"))[, -grep(glbFeatsCategory, names(ctgry_df))]))
### bid0_sp
# Better than MFO; models.n=28; min.RMSE.fit=0.0521233; err.abs.fit.sum=7.3631895
# old: Top x from auto; models.n= 5; min.RMSE.fit=0.06311047; err.abs.fit.sum=9.5937080
# RFE only ; models.n=16; min.RMSE.fit=0.05148588; err.abs.fit.sum=7.2875091
# RFE subset only ;models.n= 5; min.RMSE.fit=0.06040702; err.abs.fit.sum=9.059088
# RFE subset only ;models.n= 9; min.RMSE.fit=0.05933167; err.abs.fit.sum=8.7421288
# RFE subset only ;models.n=15; min.RMSE.fit=0.0584607; err.abs.fit.sum=8.5902066
# RFE subset only ;models.n=17; min.RMSE.fit=0.05496899; err.abs.fit.sum=8.0170431
# RFE subset only ;models.n=18; min.RMSE.fit=0.05441577; err.abs.fit.sum=7.837223
# RFE subset only ;models.n=16; min.RMSE.fit=0.05441577; err.abs.fit.sum=7.837223
### bid0_sp
### bid1_sp
# "auto"; err.abs.fit.sum=76.699774; min.RMSE.fit=0.2186429
# "RFE.X.*"; err.abs.fit.sum=; min.RMSE.fit=0.221114
### bid1_sp
indep_vars <- paste(mygetPredictIds(glb_rsp_var)$value, glb_mdl_ensemble, sep = "")
if (glb_is_classification)
indep_vars <- paste(indep_vars, ".prob", sep = "")
# Some models in glb_mdl_ensemble might not be fitted e.g. RFE.X.Interact
indep_vars <- intersect(indep_vars, names(glbObsFit))
# indep_vars <- grep(mygetPredictIds(glb_rsp_var)$value, names(glbObsFit), fixed=TRUE, value=TRUE)
# if (glb_is_regression)
# indep_vars <- indep_vars[!grepl("(err\\.abs|accurate)$", indep_vars)]
# if (glb_is_classification && glb_is_binomial)
# indep_vars <- grep("prob$", indep_vars, value=TRUE) else
# indep_vars <- indep_vars[!grepl("err$", indep_vars)]
#rfe_fit_ens_results <- myrun_rfe(glbObsFit, indep_vars)
for (method in c("glm", "glmnet")) {
for (trainControlMethod in
c("boot", "boot632", "cv", "repeatedcv"
#, "LOOCV" # tuneLength * nrow(fitDF)
, "LGOCV", "adaptive_cv"
#, "adaptive_boot" #error: adaptive$min should be less than 3
#, "adaptive_LGOCV" #error: adaptive$min should be less than 3
)) {
#sav_models_df <- glb_models_df; all.equal(sav_models_df, glb_models_df)
#glb_models_df <- sav_models_df; print(glb_models_df$id)
if ((method == "glm") && (trainControlMethod != "repeatedcv"))
# glm used only to identify outliers
next
ret_lst <- myfit_mdl(
mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = paste0(mdl_id_pfx, ".", trainControlMethod),
type = glb_model_type, tune.df = NULL,
trainControl.method = trainControlMethod,
trainControl.number = glb_rcv_n_folds,
trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = method)),
indep_vars = indep_vars, rsp_var = glb_rsp_var,
fit_df = glbObsFit, OOB_df = glbObsOOB)
}
}
dsp_models_df <- get_dsp_models_df()
}
if (is.null(glb_sel_mdl_id))
glb_sel_mdl_id <- dsp_models_df[1, "id"] else
print(sprintf("User specified selection: %s", glb_sel_mdl_id)) ## [1] "User specified selection: All.X##rcv#glmnet"myprint_mdl(glb_sel_mdl <- glb_models_lst[[glb_sel_mdl_id]])
## Length Class Mode
## a0 59 -none- numeric
## beta 177 dgCMatrix S4
## df 59 -none- numeric
## dim 2 -none- numeric
## lambda 59 -none- numeric
## dev.ratio 59 -none- numeric
## nulldev 1 -none- numeric
## npasses 1 -none- numeric
## jerr 1 -none- numeric
## offset 1 -none- logical
## call 5 -none- call
## nobs 1 -none- numeric
## lambdaOpt 1 -none- numeric
## xNames 3 -none- character
## problemType 1 -none- character
## tuneValue 2 data.frame list
## obsLevels 1 -none- logical
## [1] "min lambda > lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243053.5207 -347.2823 1113.4773
## [1] "max lambda < lambdaOpt:"
## (Intercept) .pos BODY.WEIGHT
## 243140.3487 -349.8186 1127.9216## [1] TRUE# From here to save(), this should all be in one function
# these are executed in the same seq twice more:
# fit.data.training & predict.data.new chunks
print(sprintf("%s fit prediction diagnostics:", glb_sel_mdl_id))## [1] "All.X##rcv#glmnet fit prediction diagnostics:"glbObsFit <- glb_get_predictions(df = glbObsFit, mdl_id = glb_sel_mdl_id,
rsp_var = glb_rsp_var)
print(sprintf("%s OOB prediction diagnostics:", glb_sel_mdl_id))## [1] "All.X##rcv#glmnet OOB prediction diagnostics:"glbObsOOB <- glb_get_predictions(df = glbObsOOB, mdl_id = glb_sel_mdl_id,
rsp_var = glb_rsp_var)
glb_featsimp_df <-
myget_feats_importance(mdl=glb_sel_mdl, featsimp_df=NULL)
glb_featsimp_df[, paste0(glb_sel_mdl_id, ".imp")] <- glb_featsimp_df$imp
#mdl_id <-"RFE.X.glmnet"; glb_featsimp_df <- myget_feats_importance(glb_models_lst[[mdl_id]], glb_featsimp_df); glb_featsimp_df[, paste0(mdl_id, ".imp")] <- glb_featsimp_df$imp; print(glb_featsimp_df)
#print(head(sbst_featsimp_df <- subset(glb_featsimp_df, is.na(RFE.X.glmnet.imp) | (abs(RFE.X.YeoJohnson.glmnet.imp - RFE.X.glmnet.imp) > 0.0001), select=-imp)))
#print(orderBy(~ -cor.y.abs, subset(glb_feats_df, id %in% c(row.names(sbst_featsimp_df), "startprice.dcm1.is9", "D.weight.post.stop.sum"))))
print(glb_featsimp_df)## imp All.X##rcv#glmnet.imp
## BODY.WEIGHT 100.00000 100.00000
## .rnorm 23.74349 23.74349
## .pos 0.00000 0.00000# Used again in fit.data.training & predict.data.new chunks
glb_analytics_diag_plots <- function(obs_df, mdl_id, prob_threshold=NULL) {
if (!is.null(featsimp_df <- glb_featsimp_df)) {
featsimp_df$feat <- gsub("`(.*?)`", "\\1", row.names(featsimp_df))
featsimp_df$feat.interact <- gsub("(.*?):(.*)", "\\2", featsimp_df$feat)
featsimp_df$feat <- gsub("(.*?):(.*)", "\\1", featsimp_df$feat)
featsimp_df$feat.interact <-
ifelse(featsimp_df$feat.interact == featsimp_df$feat,
NA, featsimp_df$feat.interact)
featsimp_df$feat <-
gsub("(.*?)\\.fctr(.*)", "\\1\\.fctr", featsimp_df$feat)
featsimp_df$feat.interact <-
gsub("(.*?)\\.fctr(.*)", "\\1\\.fctr", featsimp_df$feat.interact)
featsimp_df <- orderBy(~ -imp.max,
summaryBy(imp ~ feat + feat.interact, data=featsimp_df,
FUN=max))
#rex_str=":(.*)"; txt_vctr=tail(featsimp_df$feat); ret_lst <- regexec(rex_str, txt_vctr); ret_lst <- regmatches(txt_vctr, ret_lst); ret_vctr <- sapply(1:length(ret_lst), function(pos_ix) ifelse(length(ret_lst[[pos_ix]]) > 0, ret_lst[[pos_ix]], "")); print(ret_vctr <- ret_vctr[ret_vctr != ""])
featsimp_df <- subset(featsimp_df, !is.na(imp.max))
if (nrow(featsimp_df) > 5) {
warning("Limiting important feature scatter plots to 5 out of ",
nrow(featsimp_df))
featsimp_df <- head(featsimp_df, 5)
}
# if (!all(is.na(featsimp_df$feat.interact)))
# stop("not implemented yet")
rsp_var_out <- mygetPredictIds(glb_rsp_var, mdl_id)$value
for (var in featsimp_df$feat) {
plot_df <- melt(obs_df, id.vars = var,
measure.vars = c(glb_rsp_var, rsp_var_out))
print(myplot_scatter(plot_df, var, "value", colorcol_name = "variable",
facet_colcol_name = "variable", jitter = TRUE) +
guides(color = FALSE))
}
}
if (glb_is_regression) {
if (is.null(featsimp_df) || (nrow(featsimp_df) == 0))
warning("No important features in glb_fin_mdl") else
print(myplot_prediction_regression(df=obs_df,
feat_x=ifelse(nrow(featsimp_df) > 1, featsimp_df$feat[2],
".rownames"),
feat_y=featsimp_df$feat[1],
rsp_var=glb_rsp_var, rsp_var_out=rsp_var_out,
id_vars=glb_id_var)
# + facet_wrap(reformulate(featsimp_df$feat[2])) # if [1 or 2] is a factor
# + geom_point(aes_string(color="<col_name>.fctr")) # to color the plot
)
}
if (glb_is_classification) {
if (is.null(featsimp_df) || (nrow(featsimp_df) == 0))
warning("No features in selected model are statistically important")
else print(myplot_prediction_classification(df = obs_df,
feat_x = ifelse(nrow(featsimp_df) > 1,
featsimp_df$feat[2], ".rownames"),
feat_y = featsimp_df$feat[1],
rsp_var = glb_rsp_var,
rsp_var_out = rsp_var_out,
id_vars = glb_id_var,
prob_threshold = prob_threshold))
}
}
if (glb_is_classification && glb_is_binomial)
glb_analytics_diag_plots(obs_df = glbObsOOB, mdl_id = glb_sel_mdl_id,
prob_threshold = glb_models_df[glb_models_df$id == glb_sel_mdl_id,
"opt.prob.threshold.OOB"]) else
glb_analytics_diag_plots(obs_df = glbObsOOB, mdl_id = glb_sel_mdl_id) 
## PTID BODY.WEIGHT HOSPITAL.COST .src .rnorm .pos HospCost.cut.fctr
## 14 7 60 887350 Test -1.2300370 14 (3e+05,9e+05]
## 4 2 41 809130 Test 0.9272271 4 (3e+05,9e+05]
## 26 13 71 711616 Test -0.8164872 26 (3e+05,9e+05]
## 2 1 49 660293 Test 1.4350423 2 (3e+05,9e+05]
## 72 36 6 551809 Test -0.1374965 72 (3e+05,9e+05]
## HOSPITAL.COST.All.X..rcv.glmnet HOSPITAL.COST.All.X..rcv.glmnet.err
## 14 305274.6 582075.4
## 4 287516.9 521613.1
## 26 313393.9 398222.1
## 2 297155.3 363137.7
## 72 224727.3 327081.7
## HOSPITAL.COST.All.X..rcv.glmnet.err.abs
## 14 582075.4
## 4 521613.1
## 26 398222.1
## 2 363137.7
## 72 327081.7
## HOSPITAL.COST.All.X..rcv.glmnet.is.acc .label
## 14 FALSE 7
## 4 FALSE 2
## 26 FALSE 13
## 2 FALSE 1
## 72 FALSE 36
if (!is.null(glbFeatsCategory)) {
glbLvlCategory <- merge(glbLvlCategory,
myget_category_stats(obs_df = glbObsFit, mdl_id = glb_sel_mdl_id,
label = "fit"),
by = glbFeatsCategory, all = TRUE)
row.names(glbLvlCategory) <- glbLvlCategory[, glbFeatsCategory]
glbLvlCategory <- merge(glbLvlCategory,
myget_category_stats(obs_df = glbObsOOB, mdl_id = glb_sel_mdl_id,
label="OOB"),
#by=glbFeatsCategory, all=TRUE) glb_ctgry-df already contains .n.OOB ?
all = TRUE)
row.names(glbLvlCategory) <- glbLvlCategory[, glbFeatsCategory]
if (any(grepl("OOB", glbMdlMetricsEval)))
print(orderBy(~-err.abs.OOB.mean, glbLvlCategory)) else
print(orderBy(~-err.abs.fit.mean, glbLvlCategory))
print(colSums(glbLvlCategory[, -grep(glbFeatsCategory, names(glbLvlCategory))]))
}## HospCost.cut.fctr .n.OOB .n.Fit .n.Tst .freqRatio.Fit
## (3e+05,9e+05] (3e+05,9e+05] 34 34 34 0.1370968
## [0,1e+05] [0,1e+05] 19 19 19 0.0766129
## (2e+05,3e+05] (2e+05,3e+05] 44 44 44 0.1774194
## (1e+05,2e+05] (1e+05,2e+05] 151 151 151 0.6088710
## .freqRatio.OOB .freqRatio.Tst err.abs.fit.sum
## (3e+05,9e+05] 0.1370968 0.1370968 5506659
## [0,1e+05] 0.0766129 0.0766129 1436089
## (2e+05,3e+05] 0.1774194 0.1774194 2541449
## (1e+05,2e+05] 0.6088710 0.6088710 8265777
## err.abs.fit.mean .n.fit err.abs.OOB.sum err.abs.OOB.mean
## (3e+05,9e+05] 161960.56 34 5518492 162308.60
## [0,1e+05] 75583.63 19 1429476 75235.58
## (2e+05,3e+05] 57760.20 44 2549106 57934.22
## (1e+05,2e+05] 54740.24 151 8243850 54595.04
## .n.OOB .n.Fit .n.Tst .freqRatio.Fit
## 248.0 248.0 248.0 1.0
## .freqRatio.OOB .freqRatio.Tst err.abs.fit.sum err.abs.fit.mean
## 1.0 1.0 17749973.7 350044.6
## .n.fit err.abs.OOB.sum err.abs.OOB.mean
## 248.0 17740924.6 350073.4write.csv(glbObsOOB[, c(glb_id_var,
grep(glb_rsp_var, names(glbObsOOB), fixed=TRUE, value=TRUE))],
paste0(gsub(".", "_", paste0(glb_out_pfx, glb_sel_mdl_id), fixed=TRUE),
"_OOBobs.csv"), row.names=FALSE)
fit.models_2_chunk_df <-
myadd_chunk(NULL, "fit.models_2_bgn", label.minor = "teardown")## label step_major step_minor label_minor bgn end elapsed
## 1 fit.models_2_bgn 1 0 teardown 51.435 NA NAglb_chunks_df <- myadd_chunk(glb_chunks_df, "fit.models", major.inc=FALSE)## label step_major step_minor label_minor bgn end elapsed
## 12 fit.models 6 2 2 44.146 51.445 7.299
## 13 fit.models 6 3 3 51.445 NA NA# if (sum(is.na(glbObsAll$D.P.http)) > 0)
# stop("fit.models_3: Why is this happening ?")
#stop(here"); glb2Sav()
sync_glb_obs_df <- function() {
# Merge or cbind ?
for (col in setdiff(names(glbObsFit), names(glbObsTrn)))
glbObsTrn[glbObsTrn$.lcn == "Fit", col] <<- glbObsFit[, col]
for (col in setdiff(names(glbObsFit), names(glbObsAll)))
glbObsAll[glbObsAll$.lcn == "Fit", col] <<- glbObsFit[, col]
if (all(is.na(glbObsNew[, glb_rsp_var])))
for (col in setdiff(names(glbObsOOB), names(glbObsTrn)))
glbObsTrn[glbObsTrn$.lcn == "OOB", col] <<- glbObsOOB[, col]
for (col in setdiff(names(glbObsOOB), names(glbObsAll)))
glbObsAll[glbObsAll$.lcn == "OOB", col] <<- glbObsOOB[, col]
}
sync_glb_obs_df()
print(setdiff(names(glbObsNew), names(glbObsAll)))## character(0)if (glb_save_envir)
save(glb_feats_df,
glbObsAll, #glbObsTrn, glbObsFit, glbObsOOB, glbObsNew,
glb_models_df, dsp_models_df, glb_models_lst, glb_sel_mdl, glb_sel_mdl_id,
glb_model_type,
file=paste0(glb_out_pfx, "selmdl_dsk.RData"))
#load(paste0(glb_out_pfx, "selmdl_dsk.RData"))
rm(ret_lst)## Warning in rm(ret_lst): object 'ret_lst' not foundreplay.petrisim(pn=glb_analytics_pn,
replay.trans=(glb_analytics_avl_objs <- c(glb_analytics_avl_objs,
"model.selected")), flip_coord=TRUE)## time trans "bgn " "fit.data.training.all " "predict.data.new " "end "
## 0.0000 multiple enabled transitions: data.training.all data.new model.selected firing: data.training.all
## 1.0000 1 2 1 0 0
## 1.0000 multiple enabled transitions: data.training.all data.new model.selected model.final data.training.all.prediction firing: data.new
## 2.0000 2 1 1 1 0
## 2.0000 multiple enabled transitions: data.training.all data.new model.selected model.final data.training.all.prediction data.new.prediction firing: model.selected
## 3.0000 3 0 2 1 0
glb_chunks_df <- myadd_chunk(glb_chunks_df, "fit.data.training", major.inc=TRUE)## label step_major step_minor label_minor bgn end
## 13 fit.models 6 3 3 51.445 55.383
## 14 fit.data.training 7 0 0 55.384 NA
## elapsed
## 13 3.939
## 14 NAStep 7.0: fit data training
#load(paste0(glb_inp_pfx, "dsk.RData"))
if (!is.null(glb_fin_mdl_id) && (glb_fin_mdl_id %in% names(glb_models_lst))) {
warning("Final model same as user selected model")
glb_fin_mdl <- glb_models_lst[[glb_fin_mdl_id]]
} else
# if (nrow(glbObsFit) + length(glbObsFitOutliers) == nrow(glbObsTrn))
if (!all(is.na(glbObsNew[, glb_rsp_var])))
{
warning("Final model same as glb_sel_mdl_id")
glb_fin_mdl_id <- paste0("Final.", glb_sel_mdl_id)
glb_fin_mdl <- glb_sel_mdl
glb_models_lst[[glb_fin_mdl_id]] <- glb_fin_mdl
} else {
if (grepl("RFE\\.X", names(glbMdlFamilies))) {
indep_vars <- myadjust_interaction_feats(subset(glb_feats_df,
!nzv & (exclude.as.feat != 1))[, "id"])
rfe_trn_results <-
myrun_rfe(glbObsTrn, indep_vars, glbRFESizes[["Final"]])
if (!isTRUE(all.equal(sort(predictors(rfe_trn_results)),
sort(predictors(rfe_fit_results))))) {
print("Diffs predictors(rfe_trn_results) vs. predictors(rfe_fit_results):")
print(setdiff(predictors(rfe_trn_results), predictors(rfe_fit_results)))
print("Diffs predictors(rfe_fit_results) vs. predictors(rfe_trn_results):")
print(setdiff(predictors(rfe_fit_results), predictors(rfe_trn_results)))
}
}
# }
if (grepl("Ensemble", glb_sel_mdl_id)) {
# Find which models are relevant
mdlimp_df <- subset(myget_feats_importance(glb_sel_mdl), imp > 5)
# Fit selected models on glbObsTrn
for (mdl_id in gsub(".prob", "",
gsub(mygetPredictIds(glb_rsp_var)$value, "", row.names(mdlimp_df), fixed = TRUE),
fixed = TRUE)) {
mdl_id_components <- unlist(strsplit(mdl_id, "[.]"))
mdlIdPfx <- paste0(c(head(mdl_id_components, -1), "Train"),
collapse = ".")
if (grepl("RFE\\.X\\.", mdlIdPfx))
mdlIndepVars <- myadjust_interaction_feats(myextract_actual_feats(
predictors(rfe_trn_results))) else
mdlIndepVars <- trim(unlist(
strsplit(glb_models_df[glb_models_df$id == mdl_id, "feats"], "[,]")))
ret_lst <-
myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = mdlIdPfx,
type = glb_model_type, tune.df = glbMdlTuneParams,
trainControl.method = "repeatedcv",
trainControl.number = glb_rcv_n_folds,
trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = tail(mdl_id_components, 1))),
indep_vars = mdlIndepVars,
rsp_var = glb_rsp_var,
fit_df = glbObsTrn, OOB_df = NULL)
glbObsTrn <- glb_get_predictions(df = glbObsTrn,
mdl_id = tail(glb_models_df$id, 1),
rsp_var = glb_rsp_var,
prob_threshold_def =
subset(glb_models_df, id == mdl_id)$opt.prob.threshold.OOB)
glbObsNew <- glb_get_predictions(df = glbObsNew,
mdl_id = tail(glb_models_df$id, 1),
rsp_var = glb_rsp_var,
prob_threshold_def =
subset(glb_models_df, id == mdl_id)$opt.prob.threshold.OOB)
}
}
# "Final" model
if ((model_method <- glb_sel_mdl$method) == "custom")
# get actual method from the mdl_id
model_method <- tail(unlist(strsplit(glb_sel_mdl_id, "[.]")), 1)
if (grepl("Ensemble", glb_sel_mdl_id)) {
# Find which models are relevant
mdlimp_df <- subset(myget_feats_importance(glb_sel_mdl), imp > 5)
if (glb_is_classification && glb_is_binomial)
indep_vars_vctr <- gsub("(.*)\\.(.*)\\.prob", "\\1\\.Train\\.\\2\\.prob",
row.names(mdlimp_df)) else
indep_vars_vctr <- gsub("(.*)\\.(.*)", "\\1\\.Train\\.\\2",
row.names(mdlimp_df))
} else
if (grepl("RFE.X", glb_sel_mdl_id, fixed = TRUE)) {
indep_vars_vctr <- myextract_actual_feats(predictors(rfe_trn_results))
} else indep_vars_vctr <-
trim(unlist(strsplit(glb_models_df[glb_models_df$id ==
glb_sel_mdl_id
, "feats"], "[,]")))
if (!is.null(glb_preproc_methods) &&
((match_pos <- regexpr(gsub(".", "\\.",
paste(glb_preproc_methods, collapse = "|"),
fixed = TRUE), glb_sel_mdl_id)) != -1))
ths_preProcess <- str_sub(glb_sel_mdl_id, match_pos,
match_pos + attr(match_pos, "match.length") - 1) else
ths_preProcess <- NULL
mdl_id_pfx <- ifelse(grepl("Ensemble", glb_sel_mdl_id),
"Final.Ensemble", "Final")
trnobs_df <- if (is.null(glbObsTrnOutliers[[mdl_id_pfx]])) glbObsTrn else
glbObsTrn[!(glbObsTrn[, glb_id_var] %in%
glbObsTrnOutliers[[mdl_id_pfx]]), ]
# Force fitting of Final.glm to identify outliers
method_vctr <- unique(c(myparseMdlId(glb_sel_mdl_id)$alg, glbMdlFamilies[["Final"]]))
for (method in method_vctr) {
#source("caret_nominalTrainWorkflow.R")
# glmnet requires at least 2 indep vars
if ((length(indep_vars_vctr) == 1) && (method %in% "glmnet"))
next
ret_lst <-
myfit_mdl(mdl_specs_lst = myinit_mdl_specs_lst(mdl_specs_lst = list(
id.prefix = mdl_id_pfx,
type = glb_model_type, trainControl.method = "repeatedcv",
trainControl.number = glb_rcv_n_folds,
trainControl.repeats = glb_rcv_n_repeats,
trainControl.classProbs = glb_is_classification,
trainControl.summaryFunction = glbMdlMetricSummaryFn,
trainControl.allowParallel = glbMdlAllowParallel,
train.metric = glbMdlMetricSummary,
train.maximize = glbMdlMetricMaximize,
train.method = method,
train.preProcess = ths_preProcess)),
indep_vars = indep_vars_vctr, rsp_var = glb_rsp_var,
fit_df = trnobs_df, OOB_df = NULL)
}
if ((length(method_vctr) == 1) || (method != "glm")) {
glb_fin_mdl <- glb_models_lst[[length(glb_models_lst)]]
glb_fin_mdl_id <- glb_models_df[length(glb_models_lst), "id"]
}
}## Warning: Final model same as glb_sel_mdl_idrm(ret_lst)## Warning in rm(ret_lst): object 'ret_lst' not foundglb_chunks_df <- myadd_chunk(glb_chunks_df, "fit.data.training", major.inc=FALSE)## label step_major step_minor label_minor bgn end
## 14 fit.data.training 7 0 0 55.384 55.815
## 15 fit.data.training 7 1 1 55.816 NA
## elapsed
## 14 0.431
## 15 NA#stop(here"); glb2Sav()
if (glb_is_classification && glb_is_binomial)
prob_threshold <- glb_models_df[glb_models_df$id == glb_sel_mdl_id,
"opt.prob.threshold.OOB"] else
prob_threshold <- NULL
if (grepl("Ensemble", glb_fin_mdl_id)) {
# Get predictions for each model in ensemble; Outliers that have been moved to OOB might not have been predicted yet
mdlEnsembleComps <- unlist(str_split(subset(glb_models_df,
id == glb_fin_mdl_id)$feats, ","))
if (glb_is_classification && glb_is_binomial)
mdlEnsembleComps <- gsub("\\.prob$", "", mdlEnsembleComps)
mdlEnsembleComps <- gsub(paste0("^",
gsub(".", "\\.", mygetPredictIds(glb_rsp_var)$value, fixed = TRUE)),
"", mdlEnsembleComps)
for (mdl_id in mdlEnsembleComps) {
glbObsTrn <- glb_get_predictions(df = glbObsTrn, mdl_id = mdl_id,
rsp_var = glb_rsp_var,
prob_threshold_def = prob_threshold)
glbObsNew <- glb_get_predictions(df = glbObsNew, mdl_id = mdl_id,
rsp_var = glb_rsp_var,
prob_threshold_def = prob_threshold)
}
}
glbObsTrn <- glb_get_predictions(df = glbObsTrn, mdl_id = glb_fin_mdl_id,
rsp_var = glb_rsp_var,
prob_threshold_def = prob_threshold)
glb_featsimp_df <- myget_feats_importance(mdl=glb_fin_mdl,
featsimp_df=glb_featsimp_df)
glb_featsimp_df[, paste0(glb_fin_mdl_id, ".imp")] <- glb_featsimp_df$imp
print(glb_featsimp_df)## All.X##rcv#glmnet.imp imp Final.All.X##rcv#glmnet.imp
## BODY.WEIGHT 100.00000 100.00000 100.00000
## .rnorm 23.74349 23.74349 23.74349
## .pos 0.00000 0.00000 0.00000if (glb_is_classification && glb_is_binomial)
glb_analytics_diag_plots(obs_df=glbObsTrn, mdl_id=glb_fin_mdl_id,
prob_threshold=glb_models_df[glb_models_df$id == glb_sel_mdl_id,
"opt.prob.threshold.OOB"]) else
glb_analytics_diag_plots(obs_df=glbObsTrn, mdl_id=glb_fin_mdl_id) 
## PTID BODY.WEIGHT HOSPITAL.COST .src .rnorm .pos HospCost.cut.fctr
## 13 7 60 887350 Train -0.4824486 13 (3e+05,9e+05]
## 3 2 41 809130 Train 2.5252625 3 (3e+05,9e+05]
## 25 13 71 711616 Train -0.6778318 25 (3e+05,9e+05]
## 1 1 49 660293 Train -0.4801124 1 (3e+05,9e+05]
## 71 36 6 551809 Train -0.2444947 71 (3e+05,9e+05]
## .lcn HOSPITAL.COST.All.X..rcv.glmnet
## 13 OOB NA
## 3 OOB NA
## 25 OOB NA
## 1 OOB NA
## 71 OOB NA
## HOSPITAL.COST.All.X..rcv.glmnet.err
## 13 NA
## 3 NA
## 25 NA
## 1 NA
## 71 NA
## HOSPITAL.COST.All.X..rcv.glmnet.err.abs
## 13 NA
## 3 NA
## 25 NA
## 1 NA
## 71 NA
## HOSPITAL.COST.All.X..rcv.glmnet.is.acc
## 13 NA
## 3 NA
## 25 NA
## 1 NA
## 71 NA
## HOSPITAL.COST.Final.All.X..rcv.glmnet
## 13 305622.6
## 3 287864.9
## 25 313741.9
## 1 297503.4
## 71 225075.4
## HOSPITAL.COST.Final.All.X..rcv.glmnet.err
## 13 581727.4
## 3 521265.1
## 25 397874.1
## 1 362789.6
## 71 326733.6
## HOSPITAL.COST.Final.All.X..rcv.glmnet.err.abs
## 13 581727.4
## 3 521265.1
## 25 397874.1
## 1 362789.6
## 71 326733.6
## HOSPITAL.COST.Final.All.X..rcv.glmnet.is.acc .label
## 13 FALSE 7
## 3 FALSE 2
## 25 FALSE 13
## 1 FALSE 1
## 71 FALSE 36
dsp_feats_vctr <- c(NULL)
for(var in grep(".imp", names(glb_feats_df), fixed=TRUE, value=TRUE))
dsp_feats_vctr <- union(dsp_feats_vctr,
glb_feats_df[!is.na(glb_feats_df[, var]), "id"])
# print(glbObsTrn[glbObsTrn$UniqueID %in% FN_OOB_ids,
# grep(glb_rsp_var, names(glbObsTrn), value=TRUE)])
print(setdiff(names(glbObsTrn), names(glbObsAll)))## [1] "HOSPITAL.COST.Final.All.X..rcv.glmnet"
## [2] "HOSPITAL.COST.Final.All.X..rcv.glmnet.err"
## [3] "HOSPITAL.COST.Final.All.X..rcv.glmnet.err.abs"
## [4] "HOSPITAL.COST.Final.All.X..rcv.glmnet.is.acc"for (col in setdiff(names(glbObsTrn), names(glbObsAll)))
# Merge or cbind ?
glbObsAll[glbObsAll$.src == "Train", col] <- glbObsTrn[, col]
print(setdiff(names(glbObsFit), names(glbObsAll)))## character(0)print(setdiff(names(glbObsOOB), names(glbObsAll)))## character(0)for (col in setdiff(names(glbObsOOB), names(glbObsAll)))
# Merge or cbind ?
glbObsAll[glbObsAll$.lcn == "OOB", col] <- glbObsOOB[, col]
print(setdiff(names(glbObsNew), names(glbObsAll)))## character(0)if (glb_save_envir)
save(glb_feats_df, glbObsAll,
#glbObsTrn, glbObsFit, glbObsOOB, glbObsNew,
glb_models_df, dsp_models_df, glb_models_lst, glb_model_type,
glb_sel_mdl, glb_sel_mdl_id,
glb_fin_mdl, glb_fin_mdl_id,
file=paste0(glb_out_pfx, "dsk.RData"))
replay.petrisim(pn=glb_analytics_pn,
replay.trans=(glb_analytics_avl_objs <- c(glb_analytics_avl_objs,
"data.training.all.prediction","model.final")), flip_coord=TRUE)## time trans "bgn " "fit.data.training.all " "predict.data.new " "end "
## 0.0000 multiple enabled transitions: data.training.all data.new model.selected firing: data.training.all
## 1.0000 1 2 1 0 0
## 1.0000 multiple enabled transitions: data.training.all data.new model.selected model.final data.training.all.prediction firing: data.new
## 2.0000 2 1 1 1 0
## 2.0000 multiple enabled transitions: data.training.all data.new model.selected model.final data.training.all.prediction data.new.prediction firing: model.selected
## 3.0000 3 0 2 1 0
## 3.0000 multiple enabled transitions: model.final data.training.all.prediction data.new.prediction firing: data.training.all.prediction
## 4.0000 5 0 1 1 1
## 4.0000 multiple enabled transitions: model.final data.training.all.prediction data.new.prediction firing: model.final
## 5.0000 4 0 0 2 1
glb_chunks_df <- myadd_chunk(glb_chunks_df, "predict.data.new", major.inc=TRUE)## label step_major step_minor label_minor bgn end
## 15 fit.data.training 7 1 1 55.816 59.331
## 16 predict.data.new 8 0 0 59.332 NA
## elapsed
## 15 3.515
## 16 NAStep 8.0: predict data new
## PTID BODY.WEIGHT HOSPITAL.COST .src .rnorm .pos HospCost.cut.fctr
## 14 7 60 887350 Test -1.2300370 14 (3e+05,9e+05]
## 4 2 41 809130 Test 0.9272271 4 (3e+05,9e+05]
## 26 13 71 711616 Test -0.8164872 26 (3e+05,9e+05]
## 2 1 49 660293 Test 1.4350423 2 (3e+05,9e+05]
## 72 36 6 551809 Test -0.1374965 72 (3e+05,9e+05]
## .lcn HOSPITAL.COST.Final.All.X..rcv.glmnet
## 14 OOB 305274.6
## 4 OOB 287516.9
## 26 OOB 313393.9
## 2 OOB 297155.3
## 72 OOB 224727.3
## HOSPITAL.COST.Final.All.X..rcv.glmnet.err
## 14 582075.4
## 4 521613.1
## 26 398222.1
## 2 363137.7
## 72 327081.7
## HOSPITAL.COST.Final.All.X..rcv.glmnet.err.abs
## 14 582075.4
## 4 521613.1
## 26 398222.1
## 2 363137.7
## 72 327081.7
## HOSPITAL.COST.Final.All.X..rcv.glmnet.is.acc .label
## 14 FALSE 7
## 4 FALSE 2
## 26 FALSE 13
## 2 FALSE 1
## 72 FALSE 36
## Loading required package: stringr## [1] "glb_sel_mdl_id: All.X##rcv#glmnet"## [1] "glb_fin_mdl_id: Final.All.X##rcv#glmnet"## [1] "Cross Validation issues:"
## MFO###lm Max.cor.Y.rcv.1X1###glmnet
## 0 0## min.RMSE.OOB max.R.sq.OOB max.Adj.R.sq.fit
## Max.cor.Y##rcv#rpart 69364.86 0.678529263 NA
## Max.cor.Y.rcv.1X1###glmnet 102101.85 0.303486397 0.2978099076
## Low.cor.X##rcv#glmnet 102280.03 0.301053181 0.2924677657
## All.X##rcv#glmnet 102280.03 0.301053181 0.2924677657
## MFO###lm 122798.07 -0.007501072 0.0007949678
## min.RMSE.fit
## Max.cor.Y##rcv#rpart 74482.18
## Max.cor.Y.rcv.1X1###glmnet 102101.17
## Low.cor.X##rcv#glmnet 102975.27
## All.X##rcv#glmnet 102975.27
## MFO###lm 122043.65## [1] "All.X##rcv#glmnet OOB RMSE: 102280.0345"## .freqRatio.Fit .freqRatio.OOB .freqRatio.Tst .n.Fit .n.OOB
## (3e+05,9e+05] 0.1370968 0.1370968 0.1370968 34 34
## [0,1e+05] 0.0766129 0.0766129 0.0766129 19 19
## (2e+05,3e+05] 0.1774194 0.1774194 0.1774194 44 44
## (1e+05,2e+05] 0.6088710 0.6088710 0.6088710 151 151
## .n.Tst .n.fit .n.new .n.trn err.abs.OOB.mean
## (3e+05,9e+05] 34 34 34 34 162308.60
## [0,1e+05] 19 19 19 19 75235.58
## (2e+05,3e+05] 44 44 44 44 57934.22
## (1e+05,2e+05] 151 151 151 151 54595.04
## err.abs.fit.mean err.abs.new.mean err.abs.trn.mean
## (3e+05,9e+05] 161960.56 162308.60 161960.56
## [0,1e+05] 75583.63 75235.58 75583.63
## (2e+05,3e+05] 57760.20 57934.22 57760.20
## (1e+05,2e+05] 54740.24 54595.04 54740.24
## err.abs.OOB.sum err.abs.fit.sum err.abs.new.sum
## (3e+05,9e+05] 5518492 5506659 5518492
## [0,1e+05] 1429476 1436089 1429476
## (2e+05,3e+05] 2549106 2541449 2549106
## (1e+05,2e+05] 8243850 8265777 8243850
## err.abs.trn.sum
## (3e+05,9e+05] 5506659
## [0,1e+05] 1436089
## (2e+05,3e+05] 2541449
## (1e+05,2e+05] 8265777
## .freqRatio.Fit .freqRatio.OOB .freqRatio.Tst .n.Fit
## 1.0 1.0 1.0 248.0
## .n.OOB .n.Tst .n.fit .n.new
## 248.0 248.0 248.0 248.0
## .n.trn err.abs.OOB.mean err.abs.fit.mean err.abs.new.mean
## 248.0 350073.4 350044.6 350073.4
## err.abs.trn.mean err.abs.OOB.sum err.abs.fit.sum err.abs.new.sum
## 350044.6 17740924.6 17749973.7 17740924.6
## err.abs.trn.sum
## 17749973.7## [1] "Final.All.X##rcv#glmnet prediction stats for glbObsNew:"
## id max.R.sq.new min.RMSE.new max.Adj.R.sq.new
## 1 All.X##rcv#glmnet 0.3010532 102280 0.2924596## All.X__rcv_glmnet.imp Final.All.X__rcv_glmnet.imp
## BODY.WEIGHT 100.00000 100.00000
## .rnorm 23.74349 23.74349## [1] "glbObsNew prediction stats:"## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
## label step_major step_minor label_minor bgn end
## 16 predict.data.new 8 0 0 59.332 68.984
## 17 display.session.info 9 0 0 68.985 NA
## elapsed
## 16 9.653
## 17 NANull Hypothesis (\(\sf{H_{0}}\)): mpg is not impacted by am_fctr.
The variance by am_fctr appears to be independent. #{r q1, cache=FALSE} # print(t.test(subset(cars_df, am_fctr == "automatic")$mpg, # subset(cars_df, am_fctr == "manual")$mpg, # var.equal=FALSE)$conf) # We reject the null hypothesis i.e. we have evidence to conclude that am_fctr impacts mpg (95% confidence). Manual transmission is better for miles per gallon versus automatic transmission.
## label step_major step_minor label_minor bgn end
## 10 fit.models 6 0 0 28.340 38.723
## 16 predict.data.new 8 0 0 59.332 68.984
## 1 import.data 1 0 0 10.550 19.452
## 12 fit.models 6 2 2 44.146 51.445
## 11 fit.models 6 1 1 38.724 44.145
## 13 fit.models 6 3 3 51.445 55.383
## 2 inspect.data 2 0 0 19.453 23.335
## 15 fit.data.training 7 1 1 55.816 59.331
## 9 select.features 5 0 0 26.580 28.339
## 5 extract.features 3 0 0 24.558 25.968
## 3 scrub.data 2 1 1 23.335 24.508
## 14 fit.data.training 7 0 0 55.384 55.815
## 6 manage.missing.data 3 1 1 25.969 26.397
## 8 partition.data.training 4 0 0 26.440 26.580
## 4 transform.data 2 2 2 24.509 24.558
## 7 cluster.data 3 2 2 26.398 26.440
## elapsed duration
## 10 10.383 10.383
## 16 9.653 9.652
## 1 8.902 8.902
## 12 7.299 7.299
## 11 5.421 5.421
## 13 3.939 3.938
## 2 3.882 3.882
## 15 3.515 3.515
## 9 1.759 1.759
## 5 1.410 1.410
## 3 1.173 1.173
## 14 0.431 0.431
## 6 0.429 0.428
## 8 0.140 0.140
## 4 0.049 0.049
## 7 0.042 0.042
## [1] "Total Elapsed Time: 68.984 secs"
## label step_major step_minor label_minor
## 3 fit.models_0_Max.cor.Y.rcv.*X* 1 2 glmnet
## 4 fit.models_0_Low.cor.X 1 3 glmnet
## 2 fit.models_0_MFO 1 1 myMFO_classfr
## 1 fit.models_0_bgn 1 0 setup
## bgn end elapsed duration
## 3 30.509 35.501 4.992 4.992
## 4 35.502 38.714 3.212 3.212
## 2 28.846 30.508 1.662 1.662
## 1 28.815 28.845 0.030 0.030
## [1] "Total Elapsed Time: 38.714 secs"